Natural language generation (NLG) is the process wherein

Reading Computer-

HENRICKSON
computers produce output in readable human languages. Such
output takes many forms, including news articles, sports reports,
prose fiction, and poetry. These computer-generated texts are
often indistinguishable from human-written texts, and they
are increasingly prevalent. NLG is here, and it is everywhere.
Generated Texts
However, readers are often unaware that what they are reading
has been computer-generated. This Element considers how
NLG conforms to and confronts traditional understandings
of authorship and what it means to be a reader. It argues that
conventional conceptions of authorship, as well as of reader
responsibility, change in instances of NLG. What is the social

Reading Computer-Generated Texts

value of a computer-generated text? What does NLG mean for
modern writing, publishing, and reading practices? Can an NLG
system be considered an author? This Element explores such
questions, while presenting a theoretical basis for future studies.

https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Cambridge Elements in Publishing and Book Culture
Series Editor:
Samantha Rayner
University College London
Associate Editor:
Leah Tether
University of Bristol

Publishing and Book Culture

Editors and Editing
ISSN 2514-8524 (online)

Leah Henrickson
ISSN 2514-8516 (print)
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press
 Elements in Publishing and Book Culture
edited by
Samantha Rayner
University College London
Leah Tether
University of Bristol

READING
COMPUTER-GENERATED
TEXTS
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Leah Henrickson
University of Leeds
 University Printing House, Cambridge CB2 8BS, United Kingdom
One Liberty Plaza, 20th Floor, New York, NY 10006, USA
477 Williamstown Road, Port Melbourne, VIC 3207, Australia
314–321, 3rd Floor, Plot 3, Splendor Forum, Jasola District Centre,
New Delhi – 110025, India
79 Anson Road, #06–04/06, Singapore 079906

Cambridge University Press is part of the University of Cambridge.

It furthers the University’s mission by disseminating knowledge in the pursuit of
education, learning, and research at the highest international levels of excellence.

www.cambridge.org
Information on this title: www.cambridge.org/9781108822862
DOI: 10.1017/9781108906463
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

© Leah Henrickson 2021

This publication is in copyright. Subject to statutory exception
and to the provisions of relevant collective licensing agreements,
no reproduction of any part may take place without the written
permission of Cambridge University Press.
First published 2021
A catalogue record for this publication is available from the British Library.
ISBN 978-1-108-82286-2 Paperback
ISSN 2514-8524 (online)
ISSN 2514-8516 (print)
Cambridge University Press has no responsibility for the persistence or accuracy of
URLs for external or third-party internet websites referred to in this publication
and does not guarantee that any content on such websites is, or will remain,
accurate or appropriate.
 Reading Computer-Generated Texts
Elements in Publishing and Book Culture

DOI: 10.1017/9781108906463
First published online: January 2021

Leah Henrickson
University of Leeds
Author for correspondence: Leah Henrickson, L.R.Henrickson@leeds.ac.uk

Abstract: Natural language generation (NLG) is the process

wherein computers produce output in readable human languages.
Such output takes many forms, including news articles, sports
reports, prose fiction, and poetry. These computer-generated texts
are often indistinguishable from human-written texts, and they are
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

increasingly prevalent. NLG is here, and it is everywhere.

However, readers are often unaware that what they are reading has
been computer-generated. This Element considers how NLG
conforms to and confronts traditional understandings of
authorship and what it means to be a reader. It argues that
conventional conceptions of authorship, as well as of reader
responsibility, change in instances of NLG. What is the social
value of a computer-generated text? What does NLG mean for
modern writing, publishing, and reading practices? Can an NLG
system be considered an author? This Element explores such
questions, while presenting a theoretical basis for future studies.

Keywords: natural language generation, computer-generated texts,

reader response, authorship, digital humanities
© Leah Henrickson 2021
ISBNs: 9781108822862 (PB), 9781108906463 (OC)
ISSNs: 2514–8524 (online), 2514–8516 (print)
 Contents

1 Introduction 1

2 Discovering Natural Language Generation 6

3 The Development of Authorship 18

4 Algorithmic Authorship and Agency 41

https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

5 Conclusion 70

References 77
 Reading Computer-Generated Texts 1

1 Introduction
Every morning, Simon walks down the street to his local café in
Wolverhampton. He orders a medium black filter coffee, which he nurses
as he flips through the Express & Star, an independent regional newspaper.
First, the cover story. Then, the less pressing items. The newspaper’s
contributors are good at what they do, tending towards fair representation
of issues and citing relevant supporting data. Today, a story entitled
‘Majority of New Mothers in Wolverhampton Are Unmarried’ catches
Simon’s eye. He reads the story’s introduction: ‘The latest figures reveal
that 56.5 per cent of the 3,476 babies born across the area in 2016 have
parents who were not married or in a civil partnership when the birth was
registered. That’s a slight increase on the previous year.’ This is a sensitive
issue, thinks Simon. Well, it takes a special kind of journalist to consider such
a subject so objectively. Simon continues reading the article, which cites
figures and statements courtesy of the Office for National Statistics. It is not
until he reaches the end of the text that Simon reads the following statement:
‘This article has been computer-generated by Urbs Media, crafting stories
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

and harnessing automation to mass localise.’

Simon looks up from the paper, quickly setting it down onto the table in
front of him. He does not know what to think about a computer-generated
news article. You read things to connect to other people’s stories, he ponders. If
a computer writes a story, you take away that human connection. But here,
there’s no criticism, there’s no opinion. I suppose it’s so factual that I don’t really
mind if this article is computer-generated or human-generated. Simon still does
not know how he feels about computer-generated texts. The only certainty:
Now I feel like I’m on the set of a sci-fi film.
This anecdote is fictitious but the technology is most certainly not.
The ‘unmarried mothers’ article (Anon., 2017) appeared on the Express &
Star’s website on 29 November 2017, as countless more computer-
generated articles appeared in other news outlets worldwide that
same day. However, most of these articles (including those by Urbs
Media) actually do not include any disclaimers about their production
processes. ‘Crafting stories and harnessing automation to mass localise’ is
a slogan that Urbs uses to describe its service on the company’s website
 2 Publishing and Book Culture

(Urbs Media, n.d.), but it does not accompany any of their articles.
Moreover, Simon’s conflicted feelings have been paraphrased from
a series of focus groups conducted to discern readers’ emotional responses
to the idea of computer-generated texts (Henrickson, 2019c). While some
may regard a computer’s ability to generate intelligible narratives as being
contained within the realm of science fiction, the technology that enables
a computer to generate cogent prose has been in development for more
than half a century. Computer scientists have long been engaged with
programming computers to generate texts that are indistinguishable from
those written by humans. Now we have reached a point when there are
systems generating texts that we may read as part of our daily routines,
unaware of their being computer-generated. The production of data-
driven sport, weather, election, and business intelligence reports has
been assigned to computers capable of producing these texts at a rate
incomparable to that of humans, and on personalised scales that could
hardly be considered efficient uses of time for paid human labour. Yet
when we read these texts, we assume that they are social products, the
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

results of human thought and intention, rather than computer-generated.

This is natural language generation (NLG). NLG is the computational
production of textual output in everyday human languages. NLG systems
are increasingly prevalent in our modern digital climate, as we have seen the
emergence of numerous companies that specialise in generating output
intended for mass readerships and readerships of one alike. Yseop works
with an international insurance company to provide personalised explana-
tions of refusals or partial agreements of credit insurance allocations
(Anon., 2019). Phrasee generates marketing copy aligned with its many
partners’ unique brand voices (Malm, 2020). Narrative Science has worked
in partnership with Deloitte to generate client-friendly narrative reports
related to such issues as budget optimisation, financial operations, and
internal auditing (Krittman, Matthews, and Glascott, 2015). These are
only a few examples of NLG’s current applications as implemented by
a burgeoning industry based around this technology. Anyone who reads
data-driven news has likely encountered at least one computer-generated
text, perhaps unknowingly. Computer-generated texts are both online and
in print, and they are everywhere.
 Reading Computer-Generated Texts 3

This Element offers an introduction to the social and literary implica-

tions of NLG from a humanities perspective. More specifically, it examines
how computer-generated texts challenge conventional understandings of
authorship and what it means to be a reader. Accentuating the humanities
perspective employed, this Element refers to NLG as ‘algorithmic author-
ship’, save for in discussions of NLG’s technical functionality and in
quotations. It should be noted that ‘algorithmic authorship’ is not a term
coined for this Element. Some (generally popular) articles have used the
term, but it has largely gone unused by NLG developers. This lack of
widespread use may reflect the current emphasis on system development
over output reception, for the term draws attention to the nuances of
authorship as it applies to human writers and NLG systems. The use of
‘algorithmic authorship’ therefore distinguishes this Element’s sociological
focus from a process-focused computer science perspective: NLG is the
technology; algorithmic authorship is the concept. While NLG refers to the
process of text production, ‘algorithmic authorship’ more clearly refers to
the social players – and the relevant sociocultural circumstances – involved
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

in this technology’s applications. ‘Algorithmic authorship’ situates NLG

systems within the complex history of authorship – a history driven by both
human and non-human agents – that is explored here.
NLG systems produce a wide range of texts for pragmatic and aesthetic
purposes. News items, image and video descriptions, and prose fiction are
just some examples of NLG’s current applications. This Element focuses
primarily on English-language prose texts for human readers, including
works that are expository (e.g. news and business reports) and aesthetic
(e.g. stories). The computer-generated texts cited in this Element exist in
static and linear reading forms. Interactive applications of NLG – for
example, chatbots – warrant their own analyses, and are thus only referred
to when contextually necessary. Likewise, those computer-generated texts
produced for reading by machines rather than humans – for example,
content generated to promote Web traffic through search engine optimisa-
tion – are outside of this Element’s scope, but have already been investi-
gated elsewhere. Further, I do not deny the importance of genre differences
in literary criticism, or wish to discredit readers’ genre-specific expectations;
expository and aesthetic texts prompt different reading practices and
 4 Publishing and Book Culture

interpretive processes, as do texts of different genres. However, the focus of

this Element is not one of literary analysis, but of readers’ responses to the
very concept of NLG itself. Once the stage has been set by such general
research, future studies are better positioned to conduct genre-specific
analyses. For now, concentration on textual output actually detracts from
the more fundamental issue of determining NLG’s unique contributions to
the modern textual landscape: a landscape permeated with varied modes of
human–computer collaboration. Without recognition of these contribu-
tions, one cannot fully appreciate the roles of NLG systems as cultural
artefacts, as reflections of contemporary social values and needs, and in turn
as re-enforcers or interrogators of these values and needs.
Of course, an NLG system’s output must at least somewhat conform to
readers’ expectations of literary convention for that output and system to be
relevant. Complete narratives comprise a beginning, a middle, and an end;
genre conventions are anticipated; every text has an author. Indeed, it is not
unreasonable to assume that a text reflects human agency, and that a text –
regardless of genre – is an effort to communicate a predetermined message.
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

With this assumption, readers engage their interpretive faculties to assign

authorial intention, developing a perceived contract with the author. I refer
to this author–reader contract as ‘the hermeneutic contract’ (Henrickson,
2018a). The establishment of an author–reader contract is warranted given
the underlying assumption that through language we articulate and legit-
imise lived experiences to ourselves and others. Our perceptions of the
world, and others’ perceptions of ourselves, are shaped and shared by the
words at our disposal. It is through language that agency is exercised, and
that potentialities are realised. The hermeneutic contract is therefore rooted
in an expectation of agency informed by lived experience.
Computer-generated texts in their current state complicate the herme-
neutic contract. The hermeneutic contract rests on two assumptions: that
readers believe that authors want readers to be interested in their texts, and
that authors want readers to understand their texts. Even in instances of the
avant-garde, authors are still presumed to have produced their texts with
some sort of communicative intention that justifies unintelligibility;
a reader’s willingness to accept such writing as intentionally incomprehen-
sible is for that reader to fulfil the hermeneutic contract. Ultimately, the
 Reading Computer-Generated Texts 5

hermeneutic contract supposes intelligibility for hypothetical readers who

are able to engage with the social and literary conventions employed by
what Tereza Pavlíčková (2013) calls the ‘imagined author’, in a nod to
Wayne Booth’s (1961) ‘implied author’. The imagined author matters,
Pavlíčková explains in a later paper with Ranjana Das (Das and
Pavlíčková, 2013: 393), because its presence contributes to a reader’s sense
of familiarity with the writing subject, in turn helping to determine the
reader’s trust in a text’s source. But what happens when the human author
appears to be removed, and agency and intention may not be readily
identifiable? The author of a computer-generated text is often an obscured
figure, an uncertain entanglement of human and computer. With an eye to
cultural vacillation, this Element offers a unique interdisciplinary consid-
eration of a technology – NLG – that has been under development for more
than half a century, but which has not yet been subject to any substantial
analysis from a humanities perspective. Drawing from book history, media
studies, computer science, digital humanities, and the social sciences, this
Element offers a peek into the kaleidoscopic state of the art.
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Section 2 – ‘Discovering Natural Language Generation’ – presents

a technical introduction to NLG, defining technical terms as they are used
within this Element: algorithm, program, system, and machine learning. It
reviews the functionalities of some historical and modern NLG systems, as
well as the history of systematised writing more generally. This section
provides the necessary context for determining whether readers’ percep-
tions of computational capability and expectations of system functionality
are justified. Section 3 – ‘The Development of Authorship’ – offers
a theoretical examination into which aspects of modern social life algorith-
mic authorship manifest and mobilise, arguing that algorithmic authorship
reflects a current tendency towards individualisation. Section 4 –
‘Algorithmic Authorship and Agency’ – suggests a semantic shift from
considering NLG systems as tools for actualising human ideation to NLG
systems as social agents – and authors – in themselves. Such a semantic shift
permits a more reflective discussion about the transformative power of these
systems’ output, explicitly recognising their distinct social contributions.
Having presented a framework for reading computer-generated texts, this
Element concludes with recommendations for further studies.
 6 Publishing and Book Culture

2 Discovering Natural Language Generation

2.1 Introduction
The year is 1845. Spectators gather in Piccadilly’s Egyptian Hall, eagerly
anticipating a demonstration of John Clark’s Eureka machine. With the
mere pull of a lever, the Eureka generates a grammatically and metrically
correct line of Latin dactylic hexameter. For the price of one shilling,
visitors witness the mechanised spectacle of the wooden, bureau-like con-
traption moving its wooden staves, metal wires, and revolving drums to
produce a line of Latin verse that appears in the machine’s front window.
While each line adheres to a predefined system of scansion and strict
syntactic formula – adjective | noun | adverb | verb | noun | adjective –
the Eureka can churn out an estimated 26 million permutations (Hall, 2007:
227). Eureka! Mechanised text production is in the spotlight.
Not everyone was excited. An item in the ‘Miscellanea’ section of The
Athenæum’s June 1845 issue (Anon., 1845) notes that Clark had worked on
the machine for thirteen years ‘as it would seem from the mere sport of the
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

thing, and in a spirit of indifference as to what might be its subsequent

use. . . . I do not see its immediate utility.’ Other contemporary critics too
expressed scepticism: one reviewer (Anon., 1846: 133) deemed the machine
a ‘useless toy’, while another (Nuttall, 1845: 140) asserted that it was ‘little
better than a mere puzzle, which any school-boy might perform by a simpler
process.’ Modern scholar Jason David Hall (2007: 228), however, has
described the machine as ‘much more than a show-place diversion: this
kitsch device – at once the technological embodiment of and a parody of
Victorian prosodic science [the study of verse and meter] – was a literally
interactive discursive site, the focus of a popular prosodic discourse that
existed alongside institutional debates.’ Hall sees the Eureka as a meeting
place between those of academe and members of the general public (who
could spare the shilling required to enter the exhibition). Indeed, the
machine was a ‘material embodiment . . . of the institutional practice of
Victorian prosody’ that spoke to academics and amateurs alike (Hall, 2007:
234). According to Hall, the Eureka evoked curiosity in various publics at
the onset of the golden age of automata: machines programmed to operate
 Reading Computer-Generated Texts 7

on their own, often imitating the actions of humans or animals. Mechanised

writing seemed especially enticing, with some American and British news-
papers reporting on (fictitious) writing machines that tended towards
producing poetry (Anon., 1841a; Anon., 1841b; Anon., 1844).1 The
Eureka serves as an early tangible example of mechanised writing’s long-
standing appeal to mass curiosity.
Technically, the Eureka was primitive; it just randomly applied words in
accordance with strict programmed constraints. ‘All it really did,’ Jason
David Hall (2017: 130) explains, ‘was combine a limited array of integers,
uniting predictability (all values were determined beforehand) with unpre-
dictability (the order in which the values might be arranged was not).’
However, the enthusiasm garnered for the Eureka’s mechanised production
of verse speaks not just to the Victorians’ fascination with automata, but also
to a continued enchantment with technological capability. The Eureka
seems more a mechanical toy than any convincing form of autonomy,
a bibelot rather than a bard. Nevertheless, recent efforts to restore the
machine (Eureka AHRC Project, n.d.) perpetuate the sense of curiosity
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

about a technology that helped pave the way for what we now call NLG.
NLG enjoyed a surge of interest within the academic community
throughout the 1970s to 1990s. This interest also permeated the popular
sphere. In 1986, Arthur C. Clarke (2000) published a fictional short story
about ‘The Steam-Powered Word Processor’, which Reverend Charles
Cabbage uses to mindlessly produce his sermons. Even earlier, in 1954,
Roald Dahl wrote about ‘The Great Automatic Grammatisator’: a novel-
producing machine that one Adolph Knipe uses to render human writers
obsolete. Despite such widespread interest, though, NLG research has only
recently begun engendering distinct intellectual traditions, especially
related to developmental approaches. As David McDonald (1986: 12)
suggests, this may be a result of the individualised nature of systems.
Indeed, few NLG systems comprising the field’s lineage remain. Written
in programming languages now extinct, or saved in digital formats that have
deteriorated or disappeared altogether, many of these systems survive only
through secondary literature. Further, developers tend not to build upon

1
Thanks to James Ryan for his digital curation of these materials.
 8 Publishing and Book Culture

work that has already been done, perhaps for reasons of inadequate digital
preservation and/or claims for intellectual property. The field is in disarray,
with no comprehensive analysis of NLG output reception ever having been
published. As a result, we do not know where computer-generated texts fit
within our current conceptions of authorship and reading. This section
offers the necessary technical context for succeeding sections, which exam-
ine where computer-generated texts fit within conventional understandings
of authorship and what it means to be a reader as per the hermeneutic
contract.

2.2 Technical Terminology

Preceding any substantial analysis of NLG, numerous technological terms
must be defined. This is, as any etymologist might anticipate, not so easily
done. Despite technological terms like ‘program’ and ‘algorithm’ seeping into
common speech, the meanings of these words as characterised by historical
and current usage are hardly stable or precise. The definitions provided here
are working definitions for this Element and should not be considered
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

comprehensive or static. Technological developments force alterations of

all of these terms – alterations that are hardly consistent across fields of
study. Rather than offer conclusive definitions, the following paragraphs
simply offer semantic scrutiny that informs the following discussion.
Colloquial understandings of algorithms often synonymise them with
computer programs. However, these terms are not synonymous. Using
a programming language (e.g. Java or C++), a program executes a set of
instructions provided by a developer to solve a problem through fulfilment
of a predetermined task. Each of these instructions is an algorithm.
A program may comprise just one algorithm, or it may comprise a series
of algorithms. An algorithm is a function with at least one defined source of
input that produces a defined output. In some cases, the defined output may
be an instruction to refer to another algorithm. In all cases, algorithms must
be unambiguously specific. This is demonstrated by an activity popular with
Introduction to Computer Science teachers, during which students direct
the teacher to make a peanut butter and jelly sandwich. When a student tells
the teacher to ‘put the jelly on the bread’ (an instruction), the teacher may
then dunk her hand into the jelly jar and messily slop the jelly onto the loaf.
 Reading Computer-Generated Texts 9

The student must instead direct the teacher to ‘grip the knife’s handle with
your right hand’, and then to ‘move your right hand, still gripping the knife,
up approximately seven inches’, and so on. Each of these instructions
represents an algorithm. Recipe analogies are commonly used when clar-
ifying what an algorithm is. One extra teaspoon of baking powder can
collapse a cake.
This Element, however, refers to NLG systems rather than programs.
This is because generated output more often results from a series of
programs than from a single program. Each program informs the next
program’s functionality, and together these programs make a system.
But all aspects of a system may not remain static. There are, after all,
numerous problems that cannot be solved with fixed input data, instead
necessitating an ever-growing and/or unstable corpus. In these instances,
machine learning algorithms are more suitable. Search engines employ
learning algorithms to navigate the Web and provide more relevant results
in a digital landscape constantly altered by the daily emergence of countless
new websites. News platforms employ learning algorithms to summarise
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

current events by drawing from myriad articles and social media posts.
A machine learning system has by necessity achieved a sort of autopoiesis,
capable of maintaining itself through a network of processes that ensures the
system’s continuous production, maintenance, and improvement.
A machine learning system could be viewed as a small technological
ecosystem, with limited autonomy within its specified domain. As its
algorithms evolve in response to a sort of lived experience, the system
becomes increasingly independent of its creator.
Given the semantic, grammatical, and syntactic complexities of natural
language, inputting all of a language’s rules (and exceptions) into an
NLG system would be a daunting task for even the most skilled linguist.
What is more, by the time all the rules had been inputted, the linguistic
landscape would have changed. For these reasons, many new NLG
systems employ unsupervised machine learning, which applies inductive
logic associated with unlabelled data by using clustering and association
techniques to detect patterns that humans might overlook. Rather than
teach a system a language, a developer may instead teach the system how
to learn language, thereby allowing it to update its vocabulary, grammar,
 10 Publishing and Book Culture

and syntax according to actual contemporary usage. Tang et al. (2016: 2)

identify two approaches to NLG: rule- or template-based approaches and
machine learning approaches, with the latter now often employing recur-
rent neural networks (RNNs). Many historical systems have tended
towards the former. A recent example of the latter approach comes
from a team at Google Brain. In preliminary tests, the team’s RNN
used unsupervised machine learning to distinguish linguistic patterns
across approximately 12,000 (mostly fiction) e-books, amounting to
approximately 80 million sentences, from the English Google Books
Corpora (Bowman et al., 2016). Instead of constructing sentences one
word at a time through next-step predictions (e.g. through Markov
chains, which use statistical probabilities to generate texts word by
word), Google Brain’s system can construct sentences in more complex
ways that accommodate global concepts and ‘big idea’ conveyance,
resulting in output that better mimics human writing styles. In another
recent example, Microsoft’s Twitterbot Tay (@TayandYou), ‘a chatbot
created for 18- to 24- year-olds in the U.S. for entertainment purposes,’
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

likewise mimicked human writing styles (Lee, 2016). However, those

writing styles – applied by Twitter users interacting with Tay – led to the
bot’s spouting lewd and racially charged remarks within hours. Tay was
promptly deactivated and Microsoft issued a formal apology.
The Tay fiasco shows that, in any machine learning system, a developer
must in some way specify what the system is intended to do, and must
embed algorithmic preferences and constraints to ensure that the system
aligns with its intended purpose. ‘If the rules of legal chess are not built into
a chess-playing machine as constraints, and if the machine is given the
power to learn, it may change without notice from a chess-playing machine
into a machine doing a totally different task,’ Norbert Wiener warns in his
1950 The Human Use of Human Beings (205–6). ‘On the other hand, a chess-
playing machine with the rules built in as constraints may still be a learning
machine as to tactics and policies.’ Tay serves as just one modern affirma-
tion of the need for constraints in machine learning NLG systems. With no
moral bounds and limited programmed appreciation for the nuances of
language, Tay was unable to avoid offence. Tay’s computational capacity
for creation was insufficiently curbed.
 Reading Computer-Generated Texts 11

2.3 Historically Salient NLG Systems

Some of the earliest forms of NLG were interactive systems that depended
upon human users inputting comments in natural languages to receive
computer-generated natural language responses. While seemingly simple,
these responses stemmed from significant technological developments
within natural language processing. Such systems, however, could hardly
be considered autonomous or capable of independent communicative inten-
tion. Their output was predictable, determined by regular user feedback.
Only two conversational systems are noted here, but such systems – now
called chatbots – continue to be developed extensively today (Natale, in
press).
ELIZA was an early chatbot developed by MIT’s Joseph Weizenbaum
from 1964 to 1966. The system adopted multiple personae through various
sets of scripts that searched for keywords and then repurposed these key-
words according to programmed rules (Weizenbaum, 1966: 37–42).
DOCTOR, the system’s most well-known script, transformed ELIZA
into a Rogerian psychotherapist. In this script, ELIZA remodelled users’
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

inputted statements into open-ended questions that prompted users to input

more details that continued the conversational cycle. ELIZA’s output was
determined by user input, and the system’s perspective was tailored to the
human individual commanding the conversation. Users felt understood
because the system adopted their own language to craft its open-ended
prompts for elaboration, repeating what had been said with only slight
alterations to syntactic form. Without the user to guide the conversation,
ELIZA was rendered mute and moot.
In the wake of ELIZA’s success, Terry Winograd developed SHRDLU
between 1968 and 1970 for his doctoral thesis, which was completed within
the same MIT lab with which Joseph Weizenbaum was affiliated.
SHRDLU’s name refers to ‘ETAOIN SHRDLU’, a phrase that hot metal
type-casters could easily use to fill space in faulty lines of type because these
letters comprised the first two columns of their keyboards. Although
SHRDLU’s name nods to its linguistic ability, it did not generate language
on its own, but demonstrated natural language understanding using
a computer-simulated scene comprising a table, a ‘hand’ (like a virtual
 12 Publishing and Book Culture

claw from an arcade grabber machine), a box, and several blocks and
pyramids of different sizes and colours. It executed commands, answered
questions, and prompted users to provide further information if needed.
SHRDLU was significant not only because it could interact with its users by
processing English-language input and responding appropriately, but also
because it had a memory of sorts. For example, the system could remember
that Winograd had defined a ‘steeple’ as a structure that contains two green
cubes and a pyramid (Winograd, 1971: 52–5). SHRDLU could also cor-
rectly answer questions about actions it had executed in the past, demon-
strating a primitive – however mechanical – lived experience.
While Winograd did not explicitly intend SHRDLU to be a work of
interactive fiction, Professor of Digital Media Nick Montfort deems
SHRDLU ‘the first work with all the formal elements of interactive fiction.’
According to Montfort (2005: 85), SHRDLU ‘allowed for more interesting
potential narratives, simulated spaces, and challenges to later be integrated
with the sort of structure it exemplified. By augmenting SHRDLU’s parser
and world model (actually, a far simplified version of these) in this way,
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

interactive fiction could be fully realized.’ SHRDLU was a precursor to NLG

as we now know it: a field that still necessarily depends upon human–
computer interaction despite increasing computational autonomy. The sys-
tems described below exemplify numerous forms such interplay has taken.
Early systems like ELIZA and SHRDLU contributed to widespread
enthusiasm for computational language use that spurred developments in
NLG. Within only a few years, NLG systems that generated texts longer
than just a few lines, and with cohesive narrative elements, were steadily
being created worldwide.
Research indicates that the earliest prose generation systems emerged
throughout the 1950s and 1960s (Montfort and Fedorova, 2012: 82–3;
Roberts, 2017; Ryan, 2017). However, Sheldon Klein et al.’s automatic novel
writer (otherwise unnamed) is widely recognised as the first narrative NLG
system. This novel writer was capable of generating 2,100-word murder-
mystery stories ‘with semantic deep structure’ in less than 19 seconds. A 1973
University of Wisconsin technical report (Klein et al., 1973: 1) touts this system
as ‘part of an automated linguistic tool so powerful and of such methodological
significance that we are compelled to claim a major breakthrough in linguistic
 Reading Computer-Generated Texts 13

and computational linguistic research.’ The automatic novel writer oper-

ated under heavy constraint, though, and its output was predictable.
Reflecting the literary objectives of its developers and users, six potential
plots were predetermined by programmed templates, and a story’s char-
acter specifications were inputted by system users through numerical
valuation of specific traits. The system generated its stories as per murder
mystery scripts: preset sequences of actions that dictated characters’
interactions within a typical murder mystery story world. Like a child
learning the intricacies of fiction, the system generated only simple
sentences comprising one independent clause each, and only within the
story domain of murder mysteries. Yet the most significant result of the
system’s development was not so much its output as its influence on the
development of the most prominent early narrative NLG system: TALE-
SPIN. The automatic novel writer’s strict adherence to coded templates
would never allow for nuanced story characters and worlds. The system’s
output included ‘nothing that says what people need, nothing that says
why people make the decisions they do, nothing that explains people’s
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

reactions to one another, nothing that represents the complexities of the

physical world. The purpose of TALE-SPIN is to address exactly these
issues’ (Meehan, 1976: 9).
TALE-SPIN was developed by James Meehan from 1975 to 1977 for
a Yale doctorate. The system generated stories resembling Aesopian fables
centred on the lives of anthropomorphised animals. Meehan described TALE-
SPIN as a goal-directed problem-solving system that used virtual blueprints to
construct story worlds according to a user’s selections from lists of character
and plot specifications. Although Meehan intended for TALE-SPIN to simu-
late human thought processes, the system could hardly be said to have
possessed any capacity for self-reflexivity (Meehan, 1976: 4–5). Meehan
(1976: 107) himself described TALE-SPIN as ‘a simulator of the real world:
Turn it on and watch all the people. The purpose of the simulator is to model
rational behavior; the people in it are supposed to act like real people. To do
that, we have to model various kinds of human thinking, how people relate to
each other socially, for example.’ Yet TALE-SPIN did not create an entire
world at once, wherein characters could move freely, but constructed its
microworlds in light of the current protagonist’s goals, revealing only what
 14 Publishing and Book Culture

was necessary for the story’s completion. Moreover, TALE-SPIN generated

its stories using three modes, all of which were heavily dependent upon user
direction (Meehan, 1976: 15–17). The user and system were co-creators, as it
was the user who decided characters and relationships, as well as plot direc-
tions and details. TALE-SPIN did not simulate human thought; its stories
were manifestations of human intention. TALE-SPIN, however, used its own
language to generate stories in its own authorial voice, strengthening the
illusion of computational self-sufficiency and agency.
Succeeding systems were developed in explicit response to TALE-SPIN,
as were AUTHOR (Dehn, 1989) and MINSTREL (Turner, 1993), or at least
nodded to their distinguished ancestor, as did UNIVERSE (Lebowitz, 1985)
and GRANDMOTHER (Casebourne, 1996). Outside of the academy,
William Chamberlain and Thomas Etter developed Racter, which generated
what was advertised as ‘the first book ever written by a computer’: The
Policeman’s Beard Is Half Constructed (Racter [Chamberlain and Etter], 1984).
Collegiate or civil, all of these systems reflect individualised – and short-lived
– efforts, and represent various developmental approaches and aims. They
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

are laden with human values, and are all human-controlled, albeit within
varying contexts of design and power. All were programmed to generate
output that resembled typical forms of human-written texts and/or mimicked
human writers’ cognitive processes during text production: namely, those
processes associated with problem solving and creativity. Some systems were
developed to facilitate reflection upon human creativity through the system-
atisation of cognitive processes associated with acts of creation; others were
developed to model human cognition, revealing that human thought pro-
cesses associated with acts of writing are too complex to mechanise entirely,
but seemingly straightforward enough to justify attempts to do so anyway.
The only one of these systems developed outside of an academic context –
Racter – emerged as an exploration of novelty, an ode to literary wackiness.
Whether for experimentation or entertainment, all of these systems have
served as means of exploring the theoretical and practical facets of creativity
as it is applied to authorship of cohesive narratives. They were born out of the
interconnectedness of humans and their technologies, and driven by curiosity
about computational competence as embodied by NLG, conscious or coded.
 Reading Computer-Generated Texts 15

2.4 Systematised Writing

To meaningfully communicate requires awareness of the self, and of the self
in relation to surrounding circumstances. NLG systems cannot (yet) be said
to possess such awareness, but many have appeared to. Although some
systems have practised programmed self-evaluation that may be seen as
mimicking human writers’ non-linear writing processes, these evaluations
have only validated the originality of generated stories in relation to past
output. While developments in machine learning may lead – are leading –
to the creation of NLG systems that are able to autonomously situate
themselves within a variety of social contexts, so long as systems depend
upon human participation in text generation through, for example, deter-
mination of plot points and character profiles, these systems can never truly
practise any behaviour of thought or creativity equivalent to that of
humans.
Yet the output from these systems is not wholly distinct from human-
written texts. Like the experimental writers of the 1960s, for example, NLG
systems depend upon the systemisation of language according to predeter-
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

mined rules that reflect communicative and aesthetic intentions. As per the
Oulipo group’s extended name – Ouvroir de littérature potentielle (work-
shop of potential literature) – potentialities of language and meaning are
explored through the conscious application of constraint. An Oulipian work
is produced less from the intention to convey a particular message and more
from the intention to discover new worlds of syntactic and semantic
possibility by applying strict linguistic constraints. Jean Lescure’s (1973)
N+7 (S+7 in French) procedure, for example, replaces every noun in
a human-written text with the seventh one listed after it in a dictionary.
In this way, the human-written text is altered to create a new text using the
same syntactic formula but resulting in a different semantic meaning.
Readers are tasked with interpreting Oulipian texts that may seem non-
sensical at first glance, with translating their gibberish into significance. For
a more playful alternative, one need only look to the popular fill-in-the-
blanks word game Mad Libs. In Mad Libs, players complete a story template
that they themselves cannot read by suggesting words with specified
narrative functions (e.g. noun, verb). The story is then read aloud, with
 16 Publishing and Book Culture

hilarity usually ensuing. An NLG system operates similarly, selecting

messages to convey in its output, as well as appropriate words for conveying
these messages. Even for the generation of non-fiction reports – seemingly
objective narratives emerging from textual or numerical data – an NLG
system engages in an aesthetic exercise of word selection to ensure that its
output is both interesting and understandable. Moreover, like an Oulipian
or Mad Libs text, part of the computer-generated text’s attraction is the
novelty of its production, and its occasional deviation from expected
authorial comprehensibility.
But it is not just word choice that may be systematised; structural models
of literature are no recent development. William Wallace Cook’s 1928
Plotto, including more than 1,000 short plot outlines for aspiring authors,
is just one early example of a writing tool that operated on the basis of
systematised language and story structure. Georges Polti’s 1895 Thirty-Six
Dramatic Situations, which suggested fewer plot outlines, is an even earlier
example. Wycliffe Hill similarly schematised plots in his 1919 Ten Million
Photoplay Plots, and later claimed to have invented a ‘Plot Robot’ that
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

generated hundreds of plots across various genres of fiction (Anon., 2010).

Centuries prior, by no later than 1677, one John Peter published a book
including tables for ‘artificial versifying | a new way to make Latin verses’.
NLG systems evidently do not represent the onset of narrative system-
atisation, but rather extend a lineage of such practice into computational
contexts.

2.5 Conclusion
And here we find ourselves, following a history of mechanised text produc-
tion beginning in 1845, but still uncertain as to where this technology fits
within our modern age. Just as with the Eureka’s spectators, modern readers
are usually offered little more than an NLG system’s surface realisations.
A system’s constraints, determined by human developers, are indistinguish-
able, hidden within a black box of code that may never be shared, perhaps
due to intellectual property protection or patent limitations. Developers
may therefore capitalise on humans’ long-standing fascination with
mechanised text production, as well as on each individual user’s instinct
to fill in a text’s conceptual gaps. Development continues, but we still do not
 Reading Computer-Generated Texts 17

know where computer-generated texts fit within our current conceptions of

authorship and what it means to be a reader. We have yet to fully discover
NLG and all that it entails.
NLG systems are still far from achieving the autonomy that one assumes
of a human writer, despite developers’ efforts. Human involvement is
integral to the creation, maintenance, and operation of these systems.
However, this involvement is rarely overt. As a result, understanding
where NLG systems fit within the conventional hermeneutic contract that
drives modern reading experiences is hardly straightforward. What does
authorship as a concept mean in instances of computer-generated texts?
How is the reader expected to respond to one of these texts? Who – or
what – is the author? The next two sections consider such questions.
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press
 18 Publishing and Book Culture

3 The Development of Authorship

3.1 Introduction
In 2012, one Amazon customer (Smiggy) enthusiastically reviewed an
unusual book. ‘This is an epic tale that rises to Orwellian heights of
greatness and is certainly nothing to be sneezed at,’ the reviewer raved.
‘A truly riveting read from cover to cover with a real snorter of a twist near
the end (no spoilers). If it was possible to rate this six stars on Amazon then
I would, I highly recommend even despite me having to save up for a few
months to afford this!’ What was this book? The 2007 Report on Facial
Tissue Stock Excluding Toweling, Napkin, and Toilet Paper: World Market
Segmentation by City, by Professor Philip M. Parker, Ph.D. (2006a), avail-
able in paperback for a mere £795.
Philip Parker is not an author in the typical sense of the word. He is
a computer programmer, and The 2007 Report on Facial Tissue Stock is one
of his thousands of computer-generated books. Parker has patented a set of
algorithms to generate, market, and distribute physical codices whose texts
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

adhere to highly formulaic templates. These are often market reports and
reference materials, generated to meet the needs of underserved commu-
nities such as those with niche interests, or even those whose languages have
a limited number of native speakers. Parker’s patent application (‘Method
and Apparatus for Automated Authoring and Marketing’ (US 7266767 B2),
2006) touts his method of text production as a means for combatting the
financial and labour pressures of traditional publishing, such as the costs and
management of authors and editors. The patent application also notes the
method as preventing human error throughout production, and boasts
greater economic profitability than traditional human authorship by using
extant print-on-demand systems, and by eliminating the need to employ
humans to write highly formulaic and repetitive texts.
Parker’s generated texts demonstrate the potential to fulfil niche
demands for which mass production would not be economically viable.
Parker’s generation system has been administered through ICON Group
International (n.d.), which sells such specialised global market research
reports as the 176-page 2007–2012 Outlook for Instant Chocolate Milk,
 Reading Computer-Generated Texts 19

Weight Control Products, Whole Milk Powder, Malted Milk Powder, and
Other Dry Milk Products Shipped in Consumer Packages Weighing 3 Pounds or
Less Excluding Nonfat Dry Milk and Infants’ Formula in Japan, for £495
(Parker, 2006b). ICON’s World Outlook series is data-driven, blending
textual templates with automatically generated analytical tables visualising
industry trends. For text-heavier content, ICON has published a Webster’s
series of reference books including the 116-page Webster’s Swedish to
English Crossword Puzzles: Level 1 for £14.95 (Parker, 2007) and other
crossword books to facilitate language learning, in addition to books
comprising historical timelines and compilations of quotations, facts, and
phrases. Most of these books are not just printed on demand – they are
generated on demand.
In 2008, Parker’s method of text production was brought to public
attention through articles in news outlets like The Guardian (Abrahams,
2008) and The New York Times (Cohen, 2008). Articles about Parker’s
books have continued to be written since, each expressing the same sense of
confusion as those initial 2008 articles: ‘Meet the Robots Writing Your
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

News Articles: The Rise of Automated Journalism’, one is titled

(McGuinness, 2014); ‘Philip Parker’s Trick For Authoring Over 1 Million
Books: Don’t Write’, is another (Bosker, 2013); as is ‘If an Algorithm
Wrote This, How Would You Even Know?’ (Podolny, 2015). Yet Parker
himself, as quoted in one of these articles (Bosker, 2013), is hardly advocat-
ing the replacement of human authors by machines. ‘There are very few
implications for writers because we’re covering areas that writers don’t or
can’t cover,’ he explains. ‘All of these projects seem very diverse, but really
they’re all in same vein [sic]: we’re using automation to reach areas that
wouldn’t have been served otherwise. And it gives people the opportunity
to see or experience content they wouldn’t have been able to otherwise.’
According to Parker, his books are intended to supplement those that
humans write, and to broaden the range of content available to readers
with specialised needs and/or interests.
The review that opened this section was undoubtedly written in jest,
poking fun at just how specialised The 2007 Report on Facial Tissue Stock is as
a result of its automated processes of production. However, not all (poten-
tial) readers of these computer-generated books so readily find humour in
 20 Publishing and Book Culture

the technology. One Amazon reviewer (Downey, 2015) for Parker’s com-
puter-generated The Official Patient’s Sourcebook on Acne Rosacea (2002)
warns:

This book was written by a computer. Don’t bother with it.

See New York Times report March 8, 2015, ‘If an Algorithm
Wrote This, How Would You Even Know?’ Philip
M. Parker is the fraud who has a patented algorithmic system
that has ‘generated more than a million books, more than
100,000 of which are available on Amazon.’ Buyer beware!!
And Parker is NOT alone.

Another reviewer (Amazon Customer) of Acne Rosacea – posting

12 February 2004, prior to Parker’s wave of publicity – advises potential
buyers to instead take ‘a trip to google.com, drop the “acne” bit at [sic] from
“acne rosacea” and surf the first few sites you find. After that you will be in
front of this book, and have something else to do with your $USD 25.’ Such
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

recommended surfing is, in actuality, akin to how Parker’s system operates.

Like a human researcher, the system sifts through a range of digital sources
and consolidates information into a cohesive narrative. As indicated by the
former comment, however, some readers do not think of NLG positively.
‘This book was written by a computer’ is, after all, more a fact than
a judgement of textual quality. The former reviewer provides no explana-
tion as to why a computer-generated text should be considered not worth
reading, and depends on the review’s readers reaching similar conclusions
of their own accord. The other reviewer provides more detail, specifying
why one should avoid buying this book: the text offers only a cursory
overview of its subject. This reviewer (Amazon Customer, 2004) continues
the criticism of Acne Rosacea, protesting that ‘later they promise “a chapter
dedicated to helping you find your peer groups”. They end up only
mentioning a prescription drug page at rosacea-control.com, CureZone
and MedHelp (of which CureZone didn’t mention rosacea at all, and
MedHelp timed out). Again they never mention the NRS or the Rosacea
Support Group.’ The book, according to this reviewer, does not offer any
unique insight into its subject, and is therefore unsatisfactory due to
 Reading Computer-Generated Texts 21

insufficient content rather than its production process. The former review
emphasises production process over textual quality, exemplifying the sym-
bolic value attributed to originality expected from human authorship. The
inextricable relationship between content and form associates the printed
word with the human-authored text, and Parker’s decision to make his
generated texts available in printed form evidently affronts the extant
human-driven institutions that determine and distribute all that is fit for
print. Acne Rosacea, alongside Parker’s other generated books, challenges
understandings of the printed word as a fundamentally human artefact, an
expression of human intelligence, a product of an author with communica-
tive intention.

3.2 The Emergence of the Author

The author has not always been regarded as an individual creative genius,
as is now customary. For example, the material conditions of medieval
manuscript production, as well as the widely held attitudes towards the
function of the writer at that time, created circumstances of textual
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

authority radically different from those of the modern age. The medieval
writer was a systematic plagiarist of sorts, copying from a variety of
source texts to compose his own (Goldschmidt, 1943: 88–93; 113).
Medieval scholars showed little regard for the individual identities of
their books’ writers, focusing instead on the ancient truths the books
held. The author was not an individual creative genius, but a messenger,
a copyist with creative licence. Certainly, there were exceptions (Saenger,
1999: 137). On a societal scale, though, medieval writers produced
content according to a markedly different set of readerly expectations
for content originality.
Printing technology transformed the authorial landscape, and the work
of E. Ph. Goldschmidt (1943), Elizabeth Eisenstein (1979), and Lucien
Febvre and Henri-Jean Martin (1997) (to list only a few) all support the
theory that print facilitated a shift away from anonymity and towards the
‘cults of personality’ (Eisenstein, 1979) that have come to characterise
modern-day print culture. Of course, authorship did not develop wholly
as a response to print; modern understandings of the author as an individual
creative genius began to emerge as early as the shift from orality to literacy
 22 Publishing and Book Culture

(Havelock, 1980: 98). Further, cultural developments like the rise of liberal-
ism and increasing privatisation also contributed to the movement away
from anonymity and towards cults of personality, as well as these cults’
resultant senses of textual authority. Eisenstein observes that ‘the veritable
explosion of “creative acts” during the seventeenth century – the so-called
“century of genius” – can be explained partly by the great increase in
possible permutations and combinations of ideas’ made possible by cross-
cultural interchanges and ‘increased [textual] output directed at relatively
stable markets’: both factors promoted by the proliferation of print
(Eisenstein, 1979: 75). In such views, print both influenced and was influ-
enced by changing conceptions of individuality and authorship. Through its
gradual crystallisation, print culture supported a shift in the cultural mindset
wherein cults of personality that praised individual genius became common-
place. The technological developments related to print had social conse-
quences: namely, the veneration of the author.
The growth of the printing trade contributed to more definitive demar-
cations of occupational roles. The publisher and printer, for example,
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

became two people rather than one, each with defined duties. One of the
most renowned models delineating printing occupations is Robert
Darnton’s Communications Circuit. Darnton (1982: 68) begins his Circuit
with the Author and the Publisher, connected with a bidirectional arrow.
The Circuit then moves consecutively to Printers (with a nod to Suppliers),
then Shippers, Booksellers, and Readers (and Binders). The Circuit closes
with a tentative unidirectional connection between Readers and Author.
Contained within the Circuit are the pressures of the ‘economic and social
conjuncture’, ‘intellectual influences and publicity’, and ‘political and legal
sanctions’. In this model, Darnton presents a rigid division of labour that
emphasises a standard materiality of printed texts produced for profit. Such
roles continue to broadly characterise today’s book trade, upholding the
symbolic value attributed to books by comprising numerous gatekeeper
functions for all that is fit for print. Although Darnton designed his
Communications Circuit for his own studies of eighteenth-century
France, book historians have since applied it liberally to a range of analogue
and digital contexts from the past and present day (e.g. Ray Murray and
Squires, 2013; van der Weel, 2001).
 Reading Computer-Generated Texts 23

One notable feature of Darnton’s Circuit – perpetuated in subsequent

applications – is its focus on the occupational agents (e.g. Suppliers,
Booksellers, Binder) who execute the processes that constitute a text’s
lifespan, as opposed to the processes themselves. Therein lies an assump-
tion of individual human agency throughout each component of the
Circuit, expressed through occupation. Nevertheless, Darnton’s emphasis
actually appears to be on the technologies associated with print them-
selves, more than on the individuals engaging with those technologies.
‘Printers’ constitute a node in themselves; ‘Suppliers’ provide paper, ink,
type, and labour; a ‘Binder’ ties the entire printed work together. While
this Circuit initially appears centred on a social structure, most obviously
through the explicit appreciation of an ‘economic and social conjuncture’
in the Circuit’s centre, that social structure operates only in relation to the
technologies it depends upon. The economic and social conjuncture is
constrained by the technologies under consideration. Individuals repre-
sented by the Circuit’s components are defined by their specialist profes-
sional roles, which stem from the divisions of labour that emerge from the
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

increasing complexity of print’s technological facets and industrialised

processes.
Conventional modern understandings of authorship, as in Darnton’s
Circuit, are often of a skilled professional role operating within a greater
network of specialist roles. For this discussion, it is significant that Darnton
begins his Circuit with the singular author. The Circuit simply could not
function without the author, as the author is portrayed as the initial – and
independent – site of text production. All begins and – as per the tentative
connection that Darnton (1982: 67) has still yet to expound save for
a perfunctory ‘Authors are readers themselves’ – ends with the author. The
aforementioned ‘cults of personality’ that have come to characterise print
culture are manifest through this imagined author, the individual creative
genius without whom print culture as we know it could not exist. Never mind
that many instances of authorship are, in fact, collaborative efforts, and that
both authorship and reading are continuous processes: social functions that
inform and are informed by every aspect of an act of text-based communica-
tion. The author as represented in Darnton’s Communications Circuit – and
as commonly believed today – is the originator of a text.
 24 Publishing and Book Culture

3.3 Reader Responsibility and the Hermeneutic Contract

In the English-speaking world, a literate person’s eyes are trained to decode
the symbols of language (e.g. the alphabet, emoji) through linear motions
(left to right and top to bottom). These linear motions were, to be sure,
habitual in manuscript tradition. The printed book, though, ‘intensified
perspective and the fixed point of view,’ as Marshall McLuhan (1994: 172)
argues. Indeed, ‘[t]he new intensity of visual stress and private point of view
in the first century of printing were united to the means of self-expression
made possible by the typographic extension of man.’ The printed word is
thus not just an extension of the eye. Like the written word, it is also an
extension of the mind of an individual, or a group of individuals, translating
intangible thought into tangible form that overcomes constraints of time
and space to communicate with anyone who shares the literacy required to
read it; the printed word too extends the minds of those readers who
engage in related interpretive acts. The proliferation of print has arguably
contributed to a restructuring of human thought processes, as typographical
‘[u]niformity, predictability, consistency, and standardisation were prerequisites
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

for [the] analytical and scientific thinking’ that defines modern rationality (van
der Weel, 2011a: 85). The ingrained notion of printed texts as expository and
linear, requiring mental concentration and patience, seeps into the human
psyche more generally, contributing to what Adriaan van der Weel has dubbed
the Order of the Book.
As van der Weel (2011a: 91) observes, our modern conception of
democracy, mainly representative democracy, rests on the assumptions
that all participating individuals (1) can access informative texts and (2)
can read those texts. The Order of the Book assigns high symbolic value to
books (especially in codex form), and to the printed word specifically,
because the texts contained within these books direct the operation of social
institutions (e.g. governmental policies) and contribute to the formation of
shared social consciousness. Although much social interaction currently
occurs in digital forms, these forms are commonly regarded as ephemeral,
and important digital texts are often printed out for future reference or
safekeeping. Printed texts also perpetuate cultural and literary heritage
through a ‘pastness of the past’ (Goody and Watt, 1963: 311) established
 Reading Computer-Generated Texts 25

by an imagined antiquity that reflects contemporary ideologies informed by

social understandings of self in relation to the past, the present, and
predictions for the future (Anderson, 2006: 44–5). That one has invested
the time to write a text, and that further people have invested the time and
money to edit, print, distribute, and read that text contributes to the printed
word’s perceived significance. The printed word is always a social artefact
with communicative value. However, it is an artefact disembodied from its
creator through the mechanised and communal processes of its production.
While the printed word may appear neutral, disconnected from its
writer, any communicative act necessitates a message, a message sender,
and a message recipient. For printed texts the sender is generally considered
the author; the recipient, the reader. The reader anticipates the text as
a product of communicative intention, a conscious effort to convey
a predetermined message, an embodiment of human agency informed by
lived experience. The reader must interpret the message, whether through
conscious consideration or wilful disregard of authorial intention. Either
way, the reader embodies the text through an imagined author, and estab-
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

lishes a perceived contract with that author. I call this author–reader

relationship ‘the hermeneutic contract’. The hermeneutic contract is rooted
in an underlying assumption that through language we articulate and
legitimise lived experiences to ourselves and others. This is not to argue
for linguistic determinism, wherein a language constrains its users’ thought
processes, but for a softer stance of linguistic relativity, wherein a language
affects its users’ cognition. Our perceptions of the world, and others’
perceptions of ourselves, are shaped – but not necessarily wholly con-
strained – by the words at our disposal. It is through language that
interpersonal agency is exercised, and that potentialities are realised.
Computer-generated texts in their current state, however, bring the
hermeneutic contract into question. The hermeneutic contract’s commu-
nication principle rests on the assumption that readers believe that authors
want their texts to be both interesting and understandable. However, the
author of a computer-generated text is an uncertain figure: human, compu-
ter, some combination thereof, or non-existent. Yet our current Order of
the Book depends upon literacy as a means for full social participation; this
dependency is based on an understanding of texts as being of social –
 26 Publishing and Book Culture

interpersonal, human – value. Where, then, does algorithmic authorship fit

within the Order of the Book, a print-based culture with largely standar-
dised textual artefacts? How are we to negotiate this new relationship
between form and content, to appraise the status and value of texts to
which our conventional hermeneutic contract may no longer apply?

3.4 Textual Interpretation and Ownership

D. F. McKenzie’s notion of the ‘sociology of texts’ accentuates the com-
municative underpinnings of media and textual artefacts. The sociology of
texts emphasises texts as being produced, transmitted, and received within
unique sociocultural circumstances that inform their interpretation and use.
It holds that ‘the claim then is no longer for their [texts’] truth as one might
seek to define that by an authorial intention, but their testimony as defined
by their historical use’ (McKenzie, 1999: 29). A sociology of texts, in
essence, recognises textual artefacts as extensions of human agency that is
manifest in various forms throughout production, transmission, and recep-
tion. A textual artefact may have myriad potential meanings, all emerging
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

from personalised interactions with that artefact occurring within particular

sociocultural contexts.
The sociology of texts is complemented by other social theories of
media like the Social Construction of Technology (SCOT) framework.
The SCOT framework hinges upon the ‘interpretive flexibility’ that
encompasses each social group’s unique hermeneutic responses to an
artefact, which are informed by the group’s own perceived needs and
contextual circumstances (Pinch and Bijker, 1984: 410–29). Stanley Fish’s
(1980: 167–73) notion of ‘interpretive communities’ is an equivalent
argument in literary studies, which Simone Murray (2018b) has revisited
for her investigation into online reading communities from a book history
perspective. Distinguishing all the potential meanings bred from a text is
hardly possible, but appreciation of that which exists outside the artefact
itself shifts attention from being fixed solely on the text towards more
comprehensive understandings of that text’s diverse past and present
applications. The very concept of authorship is similarly murky, existing
within particular social situations and within complex social and economic
networks; such circumstances are largely defined by contemporary literate
 Reading Computer-Generated Texts 27

mentalities and dominant media for text production and dissemination.

The author is a culturally constructed figure, specific both within her
individuality and position in time and space. A writer may be a corporal
individual, but an author represents the social circumstances to which that
corporal writer has grown accustomed. By applying a sociology of texts’
outward gaze, we may situate today’s computer-generated texts within the
greater sociocultural contexts in which they exist: namely, those contexts
related to conceptions of authorship and the assumed relationships
between authors and readers.
In print’s infancy, the first printer to publish a text tended to hold the
exclusive privileges to copy and reprint that text in perpetuity, with authors
receiving lump sums for their manuscripts upfront. In England, this entitle-
ment was granted by the Stationers’ Company, a guild formed in 1403 that
protected the economic interests of card-carrying members of the book
trade involved in text production and dissemination. Members of the
Company were granted ownership of published texts by entering those
texts in the Company Register, which served as an antecedent to modern
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

copyright registration. When Queen Mary I granted the Stationers’

Company its Royal Charter in 1557, the Company was legally empowered
to censor texts that offended the religious, political, or social regimes of
the day. The Stationers’ Company thereby restricted the production of
a text to one printer, and ensured that texts adhered to a social standard that
did not explicitly challenge the values and ideologies of the ruling elite.
Changing cultural values, however, led to diminished power for the
Stationers’ Company. The year 1710 saw the passing of the Statute of Anne,
commonly regarded as the world’s first copyright statute (for nuance, see
Schoff Curtin, 2017). The Statute’s extended name offers insight into its
purposes: ‘An Act for the Encouragement of Learning, by Vesting the
Copies of Printed Books in the Authors or Purchasers of such Copies,
during the Times therein mentioned.’ The Statute of Anne declared that
copyright in perpetuity, as was the Stationers’ Company’s custom, made
knowledge too expensive and exclusive, as a lack of competition allowed
a text’s copyright holder to set its price at an inaccessible amount. The
Statute made it so that copyrights were only valid for either fourteen or
twenty-one years (whether the text was published after or before the
 28 Publishing and Book Culture

Statute’s enactment, respectively), and then had to be renewed or otherwise

fall into the public domain, mimicking the contemporary legal treatment of
patents. What is more, the Statute was the first act to recognise the legal
rights of authors, with the author – rather than the printer – being granted
control over the copying of their text. Where authorship and ownership
were once distinct, they were now coincident.
Despite the Statute of Anne’s proclaimed objective for the
‘Encouragement of Learning’, financial gain has always been a motivation
driving copyright, and altruism does not make money. Thus, while the
Statute of Anne was legally enacted, works remained copyrighted in
perpetuity for the printers under common law, despite some high-profile
judicial cases involving individuals citing the Statute for protection after
they had printed texts whose copyrights had expired (Deazley, 2008). When
the Statute was finally upheld in the 1774 landmark Donaldson v Becket case,
printers finally had legal affirmation of their choices to produce inexpensive
reprints of popular texts that the Statute deemed to be within the public
domain. Literacy rates began to rise at an unprecedented rate, heightening
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

demand for printed texts and subsequently increasing both the number of
titles available and the size of print runs. There were, to be sure, other
reasons for rising literacy rates: public education initiatives, increased access
to circulating libraries, the establishment of coffee house culture, and
heightened dependency upon textual means for distribution of civic notices,
to list just a few (Cowan, 2012; Houston, 1993: 374–80). The demand for
texts was both initiated and fulfilled by the production capacity of con-
temporary printing technology; the increased presence of texts heightened
the demand for more. From the interplay of technological development and
socioeconomic circumstances, mass readerships gradually emerged to con-
sume the texts of individual authors who were granted cultural authority by
virtue of their occupation. The modern conception of the author as an
individual in an occupational role was formalised, leading to a new means
for personal financial gain: literary celebrity.
The proliferation of digital media has nuanced the conventional under-
standing of the author as literary celebrity and cultural authority. Matthew
Kirschenbaum, for one, considers the traditional author role not as dis-
appearing, but as being supplemented by an ‘@uthor’ role. Writing in the
 Reading Computer-Generated Texts 29

Los Angeles Review of Books, Kirschenbaum (2015) observes ‘a landscape of

authorship and reading that is no longer confined to simple geometries and
lines of influence, and no longer served by the established critical schools.’
According to Kirschenbaum, the @uthor exists within increasingly digital
contexts that enable new forms of author–reader interaction; his article
draws most of its examples from Twitter. In Kirschenbaum’s view, such
new digital availabilities – principally social media platforms – allow for the
public performance of authorial identities in ways that appear more sponta-
neous than staged. In her recent book about ‘reading, writing, and selling
books in the Internet era’, Simone Murray (2018a: 22–52) makes a similar
argument, framing authorship in today’s digital ecology and economy as
a conscious ongoing performance. Likewise, Gillian Beer (2014: 1) writes
about the modern interplay between author and reader. Referring specifi-
cally to digital means of distribution, Beer argues that ‘[t]he link between
human person and made work has gathered intensity exactly in the face of
new technology that more and more disperses authority: not only copyright
but the intactness of text as a good are now in question. But readers flock to
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

hear, see, and meet the author.’

Kirschenbaum, Murray, and Beer all recognise modern readers’ apparent
inability to separate the work from its creator. The work and creator are
conceptually bound, as we search for the author embedded within the text to
contextualise its production. Digital developments like social media plat-
forms facilitate connections between authors and readers and, while readers
are most certainly able to interpret texts as they deem appropriate, curiosity
about a work’s creator can now be so quickly satisfied that authorial
embodiment has been reasserted as an almost inevitable part of modern
interpretive processes. This is not to say that the author’s interpretation of
a text is final. Using William Faulkner’s 1957–1958 University of Virginia
classroom conferences to illustrate the flexibility of interpretation,
Kirschenbaum (2015) writes that ‘no one accepts Faulkner’s utterances as
the ultimate authority on his work. We’re too sophisticated for that, and
Faulkner himself cautioned that his statements and self-representations
were that of a “man in motion.” That is, Faulkner the person understood
himself to be in dialogue with his own public image.’ Readers flocked to
hear, see, and meet Faulkner. Yet his authority over his own work was
 30 Publishing and Book Culture

already dispersed in 1958, and is ever more dispersed as the implications of

the digital, notably ready access to various textual interpretations, are
pronounced. Indeed, the corporal Faulkner is not an author. Rather, read-
ers’ imagined versions of Faulkner – rooted in their interpretations of
Faulkner’s texts and public persona, as well as other critics’ interpretations
of his work – constitute his authorship, and it is this imagined author with
whom readers engage in their textual relationships.
In any act of reading there is a balance struck between author and reader:
the hermeneutic contract. This balance exists regardless of genre, although
conventions associated with a text’s genre may induce slightly different
reading approaches and understandings of authorial intention. ‘The reader
resists; the reader complies; the reader identifies; the reader re-makes the
text, not only in the light of the evidence provided but also in the presence of
an unuttered speaking voice, now speaking inside us,’ Gillian Beer (2014:
5–6) writes. ‘The author as adversary equals the reader as adversary: the
two entwine and collaborate, within the reader.’ For Beer, it is the reader who
negotiates the author–reader balance. The reader controls the reading
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

experience, and the primary responsibility for textual interpretation and

the decoding of any communicated messages lies with the messages’
receiver rather than with the messages’ transmitter. Other scholars have
similarly observed reader-centrism facilitated by the searching, personalisa-
tion, and participatory opportunities afforded by digital technologies like
the Web (van der Weel, 2014: 42–3). Computer-generated texts in parti-
cular stress the responsibility of the reader as meaning-maker. After all,
there are no readings, signings, appearances, or interviews for the NLG
system. Further, the link between person and made work is often uncertain,
with an ambiguous balance between computational contributions and
human intervention. While algorithmic authorship may represent – to
repurpose Beer’s words – a ‘new technology that more and more disperses
authority’, readers nevertheless continue seeking to embody the familiar
figure of the author, tethered to the understanding of a text as a means for
interpersonal communication. In his 1995 ‘manifesto for computer litera-
ture’, Jean-Pierre Balpe (30 [translation my own]) considers the implica-
tions of computer-generated texts for conventional understandings of
authorship. For computer-generated texts ‘more than anywhere else,’ he
 Reading Computer-Generated Texts 31

argues, ‘the author would be a fiction of the texts. This will be our final
point, but it will remain unresolved.’ In Balpe’s view, the computer-
generated text ‘destabilises’ the reader because its words ‘are emitted by
an inaccessible, superiorly reassuring mind’ rather than by an embodied
author with whom one could in principle debate a text’s meaning (Balpe,
1995: 29). The computer-generated text stimulates readers to consider their
acts of reading and meaning-making, to acknowledge and adapt their own
senses of reader responsibility. The reader’s relationship with the author –
the hermeneutic contract – is at the forefront of such consideration.

3.5 Institutions of Knowledge Dissemination and Production

Texts are artefacts separated from their creators through institutionalised
processes of production and dissemination, with the primary responsi-
bility for reconnecting artefacts and creators resting with readers.
Readerly re-embodiment of the author is the subject of Jorge Luis
Borges’ short story ‘Pierre Menard, Author of the Quixote’ (1999),
which praises the fictional Menard for his authorial genius in copying
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

verbatim two chapters of Miguel de Cervantes’ Don Quixote. As Borges’

story highlights, the perceived identity of the author is important because
answers to questions of authorship attribution reflect conceptions of both
financial and cultural capital. Authorship supports the articulation of an
individual’s voice, ownership supports capitalism’s commodification of
the individual, and together authorship and ownership (under the title of
authorship) represent a creative economy rooted in an expectation of
communicative intention compelled by individual agency. However,
digital media have confronted the more general convention of authorship
as constituting a distinct occupational role. For example, a popular desk-
top publishing software like Microsoft Word fundamentally alters one’s
writing process through its graphical user interface that allows for quick
and easy structural changes. The software may even suggest alternative
spellings or grammar usage, eliciting questions about whether a writer’s
use of desktop publishing software may be seen as a kind of co-
authorship with the computer. In a similar vein, Web-based self-
publishing platforms like Amazon’s Kindle Direct Publishing and Lulu
reduce the traditional publisher role to a series of text boxes for the
 32 Publishing and Book Culture

author to complete (see Skains, 2019 for more examples). Indeed, occu-
pational roles associated with print become blurred and altogether indis-
tinguishable in the consideration of a system that cannot easily develop
a distinct institutional order. Philip Parker’s patented algorithms are
combined to generate, market, and distribute books, but one cannot
readily ascertain precisely which part of Parker’s system does what, or
how involved Parker himself has been in the production of any one book.
In an increasingly digital world, the author continues to assert cultural
authority, but the kind of cultural authority asserted has changed.
Through digital media that permit online exchange, the author appears
more embodied than ever before, a personal figure who can respond
directly to readers through social media platforms and who may exist
outside of her own texts or even the literary realm altogether (e.g. as
a political figure or an actor). At the same time, the author – masked by
the character she performs in an appeal to literary celebrity, or perhaps
replaced by a patented set of algorithms – remains a largely imagined
figure, further disembodied from texts she has produced not only by the
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

medium of print, but by the additional production and dissemination

processes induced by digital development.
But it is the writer who determines the content of a text and, even in
instances of interactive or hypertextual prose, how that content is presented.
Adriaan van der Weel (2011a: 178–9) refers to this authority as the ‘inter-
pretive burden’, and argues that the interpretive burden is ever shifting
‘from the instigator of the communication to its recipient’ in increasingly
digital circumstances. ‘More generally, the emphasis in digital communica-
tions is shifting increasingly from the transfer of knowledge (where readers
can in principle remain passive, trusting their source) to the transfer of
information which can lead to knowledge,’ he writes. ‘The reader bears
a heavier responsibility for at least the validation of knowledge, but more
often also for its constitution.’ It is important for this discussion to distin-
guish between information and knowledge. Information refers to data that
have been systematically structured or presented in a way that connects
meaningful data points. A list of facts, for example, provides information.
Knowledge, however, refers to the critical analysis of information and the
understanding of how to apply that information meaningfully to social
 Reading Computer-Generated Texts 33

situations (Ackoff, 1989: 3–9). To one who is about to venture outdoors,

a statement like ‘the sky is grey and cloudy’ is information; that person’s
knowledge is reflected when she chooses to wear waterproof boots and
bring an umbrella. When van der Weel writes about the transfer of knowl-
edge, he is referring to the textual sharing of perceptions informed by
personal lived experiences, which may in turn be perceived by readers
altogether differently from the ways expected. Knowledge entails not only
recognition of patterns and trends within information, but also awareness of
what those patterns and trends mean within their surrounding sociocultural
circumstances. However, each reader exists within unique circumstances. It
is therefore not appropriate to adopt the formalist view of reading wherein
a reader simply follows the text, consuming content at face value. The
reader considers a text’s content in light of her own understanding of the
world, opting to in some way accept the writer’s knowledge into her own
worldview, or disregard it completely.
Many current applications of algorithmic authorship appear to tend
towards conveying information rather than knowledge. One such example
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

is Valtteri the Election Bot (Valtteri, n.d.), an NLG system developed by

a Finnish research team called Immersive Automation (see Immersive
Automation, n.d.) between December 2016 and May 2018, financially backed
by numerous public and private Finnish organisations. Valtteri populated
generic human-written English, Finnish, and Swedish templates with public
election results retrieved from the Finnish Ministry of Justice. This is a similar
production process to that of Urbs Media, described in the Introduction,
wherein humans write text-based templates with blanks for the NLG system
to populate with relevant variables for highly localised content generation.
What distinguished Valtteri’s functionality was its user-centricity. Through
Valtteri’s Web interface, users could determine which electoral district,
municipality, polling station, party, or candidate they wanted the system to
generate an article about. One article, entitled ‘The Finns Party drop most
seats across Finland’, begins with the following paragraph:

The Finns Party dropped the most council seats throughout

Finland and lost 425 seats. The party got 3.5 percentage
points fewer votes than in the last municipal election and
 34 Publishing and Book Culture

decreased their voter support by the greatest margin. The

party dropped 80501 votes since the last municipal election
and has 770 seats in the new council.

Each succeeding paragraph focuses on a different political party involved in

the April 2017 Finnish municipal elections, providing the same kind of
information in a slightly varied order. While the text’s sentences adhere to
rules of English syntax, the piece has limited linguistic flourish or variation.
Further, while the article does compare party performances to previous
elections, it offers no explicit data analysis that could aid the reader in
situating election results within historical or current societal trends. The
reader is tasked with forming knowledge from information to justify the
relevance of data to current cultural conversations. The interpretive embel-
lishments that one may expect from a human-written text are omitted, and
readers without awareness of Finnish politics are thoroughly disadvan-
taged. This is an article for those with a high level of assumed knowledge
about Finnish politics.
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

According to its developers, Valtteri was intended to serve human

journalists by identifying potentially newsworthy information, rather than
replace the human author altogether. In one case study analysis, the
developers (Leppänen et al., 2017: 181) note that ‘[t]he automation of
finding news in the era of increased availability of data sources ensures
that media companies efficiently make use of the data and report even on
the “small fish” news which they would not have had the time or
resources to report.’ For practical applications, ‘[t]his means that they
are able to cover news that is of interest to even local or small audiences,
thus improving business offering[s] and providing added value to audi-
ences.’ For Valtteri’s developers, one of the main benefits imparted by an
automated news generation system is the system’s ability to overcome the
traditional practice of gatekeeping, whereby text producers determine
what content readers should be shown. With a system like Valtteri, the
reader may freely commission articles that reflect her own personal
interests. As Valtteri’s developers (Leppänen et al., 2017: 182) claim,
‘[t]he system can produce as many news articles as possible and put the
power of selection in the reader’s hands.’ Gatekeeping, however, is an
 Reading Computer-Generated Texts 35

integral feature of the Order of the Book. The investment of numerous

individuals’ time and money in a text’s production contributes to the
value attributed to that text. The academic practice of peer review is only
one example of how such gatekeeping establishes a source as reputable
and trustworthy. Yet systems like Valtteri undermine those human-
controlled systems that exist to dictate selection and publication.
Instead, it is the reader who determines that which is fit to print.
Subsequently, it is the reader who determines what information means
in more general contexts. The reader, rather than the author, converts
information into knowledge. The interpretive burden shifts increasingly
to the reader, as the gatekeeper role appears diminished.

3.6 Hyper-Individualism
This elevated role of the reader affirms a greater societal shift towards what
I identify here as ‘hyper-individualism’. Hyper(-)individualism as a term
has already been used intermittently in discussions of psychology, media,
and culture (Gauchet, 2000; Lake, 2017). The term as it is used here, though,
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

is unconnected with these previous uses, given a lack of consistent defini-

tion. In this discussion, hyper-individualism is characterised by each per-
son’s tendency towards self-gratification, which itself is driven by senses of
self-importance and uniqueness that together manifest in a sense of entitle-
ment to highly personalised content conforming to individual worldviews.
This sense of entitlement is supported by algorithmic filters in, for example,
online search engines or social media feeds that offer customised results
based on an individual’s Web history. Individualism refers to consumers
seeking personalisation through their own curation of media. Hyper-
individualism more strongly acknowledges the pervasiveness of personali-
sation in both analogue and digital forms. Media are curated by individuals,
as well as by algorithmic filters, both conspicuous and concealed. Contrary
to the uniformity and standardisation promoted by print, algorithmic
authorship promotes mass personalisation by generating texts for reader-
ships of one.
The greater cultural discourse is shifting away from privileging the
author as an authority over a text, and fragmentedly towards each indivi-
dual reader. Indeed, digital technologies, of which NLG is only one, allow
 36 Publishing and Book Culture

for increased involvement of readers in processes of text production,

fundamentally changing the conventional relationship between author and
reader. In many instances, the author is no longer so much one who
transfers knowledge, but a co-creator with the reader. In both printed and
digital texts, the reader largely remains at the mercy of the writer’s direc-
tion, given the finite number of paths the reader may select from to journey
through the text. Nevertheless, digital texts like those produced using
Valtteri’s interactive Web interface may afford the reader a new relationship
with the text: one in which the reader assembles it, rather than watching it
unfurl. Although the number of paths the reader may choose from might be
limited, the reader’s involvement in a text’s unfurling contributes to a sense
of control over the text as one navigates through it, affirming the reader’s
senses of individual agency and importance. Of course, the argument that
reading is never linear – that readers’ eyes skip between lines of text, that
readers may not read one text from beginning to end without pause, and so
forth – has so long been commonplace that it may now be considered trite.
Few readings can ever be strictly linear, given humans’ predisposition to
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

distraction. Moreover, Web-based hypertext normalises non-linear reading

practices, as the reader must navigate through not just one text, but many.
These texts are written collaboratively, are presented in different visual
formats, may be supplemented with other media such as video or sound
clips, or may radically deviate the reader to a new subject via hyperlink. In
such modern digital contexts, it is the reader who is tasked with making
sense of this proliferation of texts, and information more generally.
Recall the distinction between information and knowledge. Information
refers to data that have been systematically structured or presented in a way
that connects meaningful points. Knowledge refers to the critical analysis of
information and the understanding of how to use that information mean-
ingfully. In the conventional hermeneutic contract, the author is understood
to be transferring knowledge to the reader. If a text is regarded as an
extension of an individual, it follows that what is written will reflect the
subjective experiences of the individual writing. Even in instances of the
avant-garde, seemingly unintelligible texts are seen as the result of com-
municative intention that the reader is left to decipher. It is the author who
encodes a text with meaning, and the reader who decodes that text. Yet in
 Reading Computer-Generated Texts 37

a hypertext-heavy world wherein the reader faces a bombardment of texts

seemingly less linear and stable than conventional human-written texts, the
reader must make sense of where each of these texts fits within more general
contexts (van der Weel, 2011b: 46). The critical analysis of information and
the understanding of how to use that information meaningfully is under-
taken by the reader, affirming a societal shift towards hyper-individualism
that is supported by inflated senses of self, resulting from a prioritisation of
individualistic self-expression. With ever-increasing access to vast amounts
of information in various analogue and digital forms, individual readers are
not only welcome to interpret textual material according to their personal
worldviews, but encouraged to generate knowledge themselves.
Yet individualised generation of knowledge rather than mere access to
formal knowledge contained in books has very real implications for conceptions
of citizenship and full participation in representative democracy. Writing from
the perspective of a sociologist of technology, Mike Michael (2014: 52–3) draws
attention to a shift wherein ‘[r]ather than educating the public so that they can
take up their roles as scientific citizens, they are to be consulted.’ As a result,
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

‘[c]itizenship here is no longer about having sufficient knowledge, or being

asked to deliberate over a range of preset technological options; it is about
having a say in the development of research programs, in agenda setting in
research funding, in prioritization of sociopolitical goals.’ Rather than having
one cultural authority to consult – an author, with printed formal knowledge –
we now have countless cultural authorities comprising a heterogeneous ‘public’,
many of whom are wary of the hierarchies of institutions related to commu-
nications and media. While this is by no means a new phenomenon (see, for
example, Gregory and Miller, 1998), the tendency towards populist anti-
intellectualism in modern rhetoric has been associated with recent political
happenings like the 2016 referendum for the United Kingdom’s withdrawal
from the European Union, as well as the result of the United States’ 2016
presidential election (Motta, 2018). Evidently stemming from an underlying
scepticism of expert knowledge and institutionalised hierarchy, anti-
intellectualism is the product of a hyper-individualised culture that has so deeply
internalised the values of individualistic facets of representative democracy that
the individual expects not only external validation for personal thoughts, but
also means for voicing and affirming subjective experiences and worldviews.
 38 Publishing and Book Culture

The individualistic desire for affirmation may partially explain why

companies specialising in the computer generation of data-driven news
articles and business reports have seen commercial success. Arria NLG,
Automated Insights, Narrative Science, and Phrasee are only a few of these
companies; they have all partnered with high-profile organisations like the
Associated Press and Microsoft to rapidly generate content for mass and
niche readerships alike. For a computer-generated business report, an NLG
system draws from quantified measures of productivity, processing these
measures to present data-driven assertions of each individual’s strengths
and weaknesses relative to other employees. Similarly, the functionality of
these systems may be applied to smaller-scale events not otherwise covered
by public news outlets. Narrative Science, for one, has partnered with
GameChanger – a baseball, softball, and basketball scorekeeping applica-
tion – ‘to craft newspaper quality “Game Recap Stories.” Within minutes of
the final out, a recap is instantly generated, and available online and in the
app’ (Anon., 2016b). GameChanger’s services are primarily marketed
towards followers of minor-league teams, including youth teams generally
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

excluded from media coverage due to lack of widespread interest. Narrative

Science and GameChanger’s collaboration lets parents read stories about
their children’s games – to see their children’s names in ‘newspaper quality’
articles – rather than require them to sift through less appealing numeric
databases. As any parent familiar with Little League sports reports will
know, though, there is an art to writing these texts. Players are always
portrayed positively, and games resulting in losses are always framed as
valiant battles rather than miserable struggles. The GameChanger game
recaps adhere to the expected format and language usage of the genre. With
this technology, any Little League game can be transmuted into a cohesive
narrative contributing to a sense of professionalism in the league, as well as
lasting excitement amongst fans. One recap, for example, reviews a local
baseball game played by ten- and eleven-year-olds. Despite the Oakville
team losing to Vaughan, ‘Evan T triggered Oakville A’s 11 U AAA’s
comeback,’ and ‘[a]fter pushing across two runs in the bottom of the fifth,
Oakville A’s 11 U AAA faced just a 5–3 deficit’ (Anon., 2016a).
Each of the generated GameChanger articles concludes with a disclaimer
that the article has been ‘Powered by Narrative Science and GameChanger
 Reading Computer-Generated Texts 39

Media’, but no description of the text’s production process accompanies the

article itself. The text seems to be a typical baseball report, with readers
uninformed of its generated nature unless they actively seek additional infor-
mation about Narrative Science. On a larger scale, Narrative Science has also
collaborated with the Big Ten Network to provide recaps of college sporting
events throughout the United States. These generated articles used to be
accompanied by an editor’s note at their conclusion, indicating that ‘[i]n
partnership with Narrative Science, BTN.com posts fast recaps after every
half of every Big Ten men’s basketball game. Use the links above for more info
on this specific game’ (BTN.com Staff, 2013). Earlier, the articles had been
accompanied by an even briefer note of their being ‘Powered by Narrative
Science’, with a link to Narrative Science’s website (BTN.com Staff, 2012).
More recent generated articles include no disclaimer related to text production.
Such disclaimers are so equivocal as to be almost meaningless, though,
and the reader must actively seek additional information about Narrative
Science to learn that an article has been computer-generated. In these
instances, it appears that it is not only readers who are aware of the symbolic
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

value attributed to human authorship, but also the companies who specialise
in wide-ranging applications of algorithmic authorship. In one study of
numerous international organisations that use news content generators,
researchers (Montal and Reich, 2017) observed what they called
a ‘transparency gap’: organisations avoiding full disclosure regarding the
origins of their computer-generated stories. Tiptoeing around issues of
authorship, organisations may share clues about NLG having factored
into article production, but omit any mention of algorithmic authorship
that may spur negative reactions similar to those expressed by the reviewers
of Parker’s Acne Rosacea. Despite the success of other NLG systems’
personalised outputs, companies reaping rewards from the mass production
of computer-generated articles appear attuned to the potential for reader
resistance, and therefore withhold production details.

3.7 Conclusion
Algorithmic authorship affronts the conventional author–reader relation-
ship – the hermeneutic contract – through hyper-individualistic personali-
sation of reading experiences. Such personalisation may take the form of
 40 Publishing and Book Culture

individualised interpretations of texts, or of the computational production

of texts for singular readerships. Using historical and modern references,
this section has shown that ‘author’ is hardly a static term, and that
technological and social developments have led to markedly different
judgements of textual authority. Likewise, the role of the reader has
changed over time. NLG, like other technologies associated with text
production, prompts reconsideration of textual functions and processes of
interpretation.
Where, then, does the conventional hermeneutic contract belong in
considerations of computer-generated texts? The establishment of the her-
meneutic contract rests almost entirely with the reader, who acknowledges
a relationship – however imaginary – with an author. As with human-written
texts, readers assume that computer-generated texts are produced with some
sort of communicative intention, to be both interesting and understandable to
an intended readership. The assumption of communicative intention is
reinforced when human writers appear increasingly embodied through use
of social media platforms and other digital media. At the same time, the
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

fragmentary nature of modern digital culture enables readers to become

increasingly responsible for crafting knowledge from vast amounts of infor-
mation, becoming what Mike Michael (2014: 52–3) calls ‘scientific citizens’,
who are to be encouraged to participate in policy decisions and agenda
setting. NLG amplifies such social fragmentation, as texts may not just be
interpreted by individual readers, but actually produced for them. In such
cases, the hermeneutic contract seems to maintain its relevance because the
responsibility for maintaining such a contract has little to do with who (or
what) the author actually is, and more to do with the reader’s perceptions of
the author. As with a human-written text, the author of a computer-generated
text is largely imagined by the reader. The next section ponders the language
that should be used to describe the algorithmic author.
 Reading Computer-Generated Texts 41

4 Algorithmic Authorship and Agency

4.1 Introduction
The date is 1st November 2017. The place is Amersfoort, the Netherlands.
Inside de Bibliotheek Eemland (the Eemland Library), an excited group has
gathered for the launch of the annual Nederland Leest (Netherlands Reads)
campaign. Run by De Stichting Collectieve Propaganda van het
Nederlandse Boek (the Foundation for the Collective Promotion for the
Dutch Book), an organisation that promotes Dutch literacy and literature,
Nederland Leest distributes free copies of a selected text in libraries across
the country each November, sparking discussion about the year’s selected
theme. In 2017, this theme was robotics. The selected text was Isaac
Asimov’s Ik, robot (I, Robot), a collection of short stories about human–
robot interaction. Nederland Leest’s reissued edition of Ik, robot, though,
concluded with a new story: ‘De robot van de machine is de mens’ (literally,
‘The Robot of the Machine is Human’). This story was written by popular
Dutch author Ronald Giphart – and co-authored by an NLG system called
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

the Asibot (Karsdorp et al., 2017).

Throughout the room two-foot-tall Nao robots (developed by
SoftBank Robotics) are scattered, with event attendees eagerly instructing
them to stand up and sit down, and engaging them in small talk. On the
stage stands four-foot-tall Pepper, a sleek humanoid robot (also devel-
oped by SoftBank Robotics). Upon walking onto the stage to begin the
event, master of ceremonies Diana Matroos briefly converses with
Pepper as though she were meeting a person. Throughout the pro-
gramme, the Nao robots and Pepper are used to testify to advancements
in robotics, and Pepper occasionally pipes up to contribute to the hour-
long discussion about the social value of robotics and mechanisation.
Audience members are asked whether they agree or disagree with
a series of statements related to the future of robotics: for example,
‘robots can make good art: agree or disagree?’ Voting with their
Nederland Leest paddles – with a robot’s head on one side and
Giphart’s face on the other – the audience’s opinion is split fairly evenly.
Then, the moment finally comes to officially launch the new edition of
 42 Publishing and Book Culture

Ik, robot. On the stage stands Ronald Giphart, positioned next to Arie
Rommers, user of one of the world’s most advanced prosthetic hands.
Rommers stands between Giphart and a large robot arm, which has been
placed on a platform adorned with a ceremonious red tablecloth.
Kraftwerk’s 1978 ‘The Robots’ blasts through the sound system.
Slowly, the robot arm lifts a copy of the new edition of Ik, robot, turning
to pass it to Rommers. From robot to cyborg, and then from cyborg to
human, Rommers passes the book to Giphart, who triumphantly raises it
towards the audience. The room erupts in applause.
At the event’s reception, attendees excitedly collect their free copies of
Ik, robot, flipping swiftly to the final chapter by Giphart and the Asibot.
Representatives for the Nao robots and Pepper stand with their machines,
answering questions about the technologies they are touting. The Nao
robots are used in community centres like libraries and seniors’ homes to
dispel fears of new technology; Pepper is used as a novelty in service-
based businesses and corporate events. All representatives assure atten-
dees that these robots repond to users through series of programmed
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

responses. ‘It’s just a puppet,’ one representative asserts. ‘He just does
what you tell him to do.’2

4.2 NLG Systems as Tools

Computers have long been employed as tools for creating textual artefacts. In
a practical sense, users engage with the technology according to its capacities.
Word-processing software, for example, offers immediate feedback on poten-
tial spelling and grammar errors, while a typewriter does not. The human and
the human’s writing tools co-evolve and co-exist, with this symbiosis influen-
cing processes of text production, dissemination, and reception. For aesthetic
purposes, human–computer symbiosis has led to text production through
such initiatives as the ALAMO (Atelier de Littérature Assistée par la
Mathématique et les Ordinateurs – literature workshop assisted by mathe-
matics and computers) (ALAMO, n.d.), an ongoing effort proposed by
Oulipians Paul Braffort and Jacques Roubaud in 1981 to promote

2
Anonymised Pepper representative, conversation with the author
(1 November 2017, Amersfoort).
 Reading Computer-Generated Texts 43

experimentation with literature through computerised systematisation and

controlled randomisation. In such cases, Margaret Boden (2010: 184–5)
argues that it is the computer artist (i.e. developer) who is responsible for
what input a system will respond to, how the system will respond, how
unpredictable the system’s output will be, and how transparent the system’s
functionality will be to users. Even the most unpredictable output, for Boden,
results from choices the computer artist has made. Likewise, Jay David Bolter
(1991: 187) has remarked that ‘[t]he autonomy of the computer is only
apparent. The computer seems to be a tool that one can let go of, but all
programs, including AI programs, simply defer the need to return to the
human user.’ Elsewhere, Bolter (1984: 232–3) explains that ‘[t]he computer is
in some ways a grand machine in the Western mechanical-dynamic tradition
and in other ways a tool-in-hand from the ancient craft tradition. The best
way to encourage the humane use of computers is to emphasize, where
possible, the second heritage over the first, the tool over the machine.’
Although a developer may not be able to predict a system’s exact output,
the output reflects choices the developer has made while programming, as
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

well as choices the user has made while engaging with the system.
One recent work that exemplifies the NLG system as tool is Zach
Whalen’s The Several Houses of Brian, Spencer, Liam, Victoria, Brayden,
Vincent, and Alex, a product of the 2017 National Novel Generation Month
(NaNoGenMo). Initiated in 2013 by Darius Kazemi (see Kazemi, n.d.),
NaNoGenMo is a spin-off of the popular National Novel Writing Month
(NaNoWriMo), when participants are challenged to write a novel of 50,000
words or more each November. In NaNoGenMo, though, participants do
not write their novels themselves. Instead, they must write codes capable of
generating 50,000-word novels. While NaNoWriMo defines a novel as an
extended piece of fiction, Kazemi explains (see nanogenmo.github.io) that
in NaNoGenMo ‘[t]he “novel” is defined however you want. It could be
50,000 repetitions of the word “meow”. It could literally grab a random
novel from Project Gutenberg. It doesn’t matter, as long as it’s 50k+
words.’ Since its inception, NaNoGenMo participants have submitted
hundreds of computer-generated novels reflecting a wide range of subjects
and generation methods. In 2014, one participant (hugovk, 2014) generated
50,000 Meows, which changed all of the words in classic works such as
 44 Publishing and Book Culture

Melville’s Moby Dick into meows of the same length; ‘Better sleep with
a sober cannibal than a drunken Christian’ translated to ‘Meeoow meoow
meow m meeow meooooow meow m meeeeow Meoooooow.’ In 2015,
another participant (Regan) wrote a code that identified the nearest named
colour for every pixel of a digital photograph of a cover image for
Hemingway’s The Sun Also Rises; the novel, totalling more than 803,000
words, comprises a list of these colours beginning with ‘Quartz. Davy’s
grey. Purple taupe. Gray.’ These kinds of works exemplify the unique
potentialities of NLG as a means for text production. Whalen’s Several
Houses takes a more traditional form, mimicking a typical children’s book,
including large, colourful text and striking imagery. Each ‘house’ is a story
that adopts the form of the classic nursery rhyme ‘This Is the House That
Jack Built’. Page spreads are split into two parts: illustrations on the left, text
on the right. ‘This is the SLEEP that eluded the WORRIED PERSON that
lay in the BED that rested the PERSON that armed the WEAPON that
hurt the PERSON that headed for the HOUSE that Spencer built,’ reads
page 128, which is accompanied by an image of a house blurred by pastel
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

watercolours and overlaid with images like a person, a gun, and a bed
(Whalen, 2018b).
To generate Several Houses, Whalen wrote a program that builds chains
of textual concepts ending in ‘house’. These chains are developed by
referencing ConceptNet, a free online semantic network that links meanings
of words to teach computers the nuances of language and intricacies of
human knowledge using machine learning methods. The accompanying
illustrations likewise stemmed from Whalen’s code. Another program
selects a relevant icon from the Noun Project’s free online repository of
icons, colours that icon, and then randomly places it on the page on top of
another Creative Commons image of a house, which itself has been
randomly selected from Flickr and coloured using a ‘watercolor’ function
on free image-editing software ImageMagick (Whalen, 2018a). As per
NaNoGenMo’s rules, Whalen has publicly shared the source codes for
both the text and the illustrations; more tech-savvy readers may generate
novels in this style for themselves (Whalen, 2018b and 2019).
Whalen’s Several Houses demonstrates the human–computer interplay
that characterises most NaNoGenMo output. It is Whalen’s – not the
 Reading Computer-Generated Texts 45

system’s – creative vision that is manifest here. Although the exact output of
the system cannot be predicted prior to generation, contributing to a sense
of computational originality, the system is a tool for realising Whalen’s
predetermined aesthetic: a humorous (and surprisingly adult-themed) take
on a classic children’s story. Really, the system produces a textual artefact
almost indistinguishable from a human-written book, capitalising upon the
conventional codex form (albeit digitised) and engrained linear reading
practices. Through Several Houses’ familiar format, Whalen subtly subverts
the textual and visual genre conventions of children’s literature, both by
satirising a classic nursery rhyme and by quantifying the artistic process so
that it is almost entirely randomised. The reader, recognising Whalen’s play
on conventional children’s literature, may appreciate the hilarity of the text
itself, but may also appreciate the perceived intention that drove the work’s
creation more generally, as well as the skill involved in coding the system.
For her 2015 NaNoGenMo submission entitled Our Arrival: A Novel,
Allison Parrish generated a diary of an expedition through fantastical
fictional places. According to the novel’s preface (written by Parrish),
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Parrish’s source corpus comprised more than 5,700 sentences from public
domain Project Gutenberg books whose subject entries included strings
related to the natural sciences, exploration, and science fiction, and whose
constructions and content satisfied Parrish’s programmed criteria (Parrish,
2015: iii). By parsing and combining randomly selected sentences according
to their grammatical constituents, Parrish’s system produced
a phantasmagorical text that invites readers on a magic carpet ride through
the outskirts of the imagination. One entry (Parrish, 2015: 223) reads:

Large fissures were well wooded as were also those behind.

They were bent and twisted out of shape.
You perceived a cradle. It was on the eve of war.
Discreetly we elected to continue. The roof stood alone in
the clearing and listened to the diminishing sounds. It had
the quality of mountains, canyons, and crags.

Although Our Arrival is formatted like a typical novel, it is linguistically and

topically outside the typical novel’s remit. A roof, after all, cannot stand,
 46 Publishing and Book Culture

nor can it listen. And it is remarkably difficult to imagine a roof with the
simultaneous qualities of mountains, canyons, and crags. Like Whalen’s
Several Houses, Our Arrival subverts the textual conventions of mass-
marketed fiction. Syntactically and semantically, it is a textual absurdity
that draws one’s attention to the potentialities of the linguistic imagination,
pushing the boundaries of the usual by repurposing extant texts in the
Project Gutenberg corpus. Moreover, both works draw attention to the
medium of the book itself. Despite using digital technology not bound to
the conventional codex format, both Whalen and Parrish have chosen to
present their computer-generated texts in familiar codex forms. Such texts
could be seen as bridging a gap between the analogue and the digital, the old
and the new, the mundane and the uncharted. Existing within systemic
structures of literary convention, readers are guided towards the interpre-
tive processes with which they have grown accustomed through cues of
familiar page layouts and intelligible texts.
Despite their similar appearances, though, human-written and compu-
ter-generated texts reflect different processes of production. Rather than
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

represent a transition from human-written to computer-generated texts,

both Whalen’s and Parrish’s work testifies to a new kind of text production
wherein NLG plays a vital role. Indeed, aesthetic endeavours capitalising
on human–computer interplay for text production – Several Houses and Our
Arrival being just two examples – seem a natural step in the evolution of
writing technologies, but hardly leave humans obsolete. The computer’s
involvement need not be limited to wiggly underlines suggesting spelling or
grammar changes, but may include participation in the selection of words
themselves, influencing – or even determining – the direction of a story.
Even in the realm of non-fictional news, this principle holds: to populate
human-written templates through programmed instruction is to partake in
the process of content curation. The computer is thus an instrument that
extends, and fundamentally transforms, the human writing process, though
it is still the human developer’s vision manifest by means of strict instruc-
tions (i.e. algorithms) resulting in profoundly constrained output.
Such output constitutes what one research group from the Dutch
Meertens Instituut and the Antwerp Centre for Digital Humanities and
Literary Criticism calls ‘synthetic literature’, resulting from ‘co-creativity’
 Reading Computer-Generated Texts 47

(Manjavacas et al., 2017). In 2017, this group launched the Asibot. Using
a graphical user interface, a human author drafts a Dutch-language text,
sentence by sentence, with the Asibot proposing sentences to continue the
story. To learn language and sentence structure, the system was trained
using the texts of 10,000 Dutch e-books. The method of machine learning
employed was largely unsupervised, meaning that the system was pro-
grammed with basic awareness of global narrative structure but otherwise
autonomously distinguished syntactic and semantic patterns within the
source corpus. The system also learned how to mimic the unique writing
styles of such renowned writers as Isaac Asimov and Dutch novelist Ronald
Giphart, and can now generate sentences using similar words and syntax as
these writers. Through machine learning the Asibot cultivated its own lived
experience of sorts: an experience that, while directed by human instruction,
stemmed from particular methods of reading and analysing a body of texts
too extensive for any one human to read in a lifetime. This experience
enables the Asibot to generate text that adheres to syntactic, semantic, and
stylistic convention, while at the same time being sufficiently original.
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Giphart himself trialled the Asibot as a co-author for his commission to

produce a new story for a Dutch edition of Isaac Asimov’s Ik, robot that was
released specially for the 2017 Nederland Leest campaign. As Giphart typed
his own sentences into the interface, the system responded by suggesting
numerous sentences to progress the story. Giphart could select one of these
suggestions, or ignore the suggestions and continue writing himself. The
system’s purpose, its developers (Manjavacas et al., 2017: 35) explain, was to
‘provoke the human writer in the process of writing.’ Giphart himself
declared that he still considered himself ‘de baas, maar hij doet het werk
[the boss, but he (the system) does the work]’ (Koolhof, 2017). The system
follows its user’s direction, but responds by exercising its own programmed
creative capacity, derived from its vast knowledge bank amassed through
machine learning. Thus, while the other above-mentioned systems were
quite clearly tools for actualising individual humans’ artistic visions, the
Asibot stands in greyer territory. This system can, in a sense, articulate its
own lived experiences by drawing from the vast corpora of texts upon
which it was trained, its use of language at least somewhat constituting
a computational equivalent of cognition. Its contributions may supplement
 48 Publishing and Book Culture

those of the human writer, revealing new avenues for literary traipsing.
Alternatively, a human user could choose to accept all of the system’s
proposals, allowing the system to produce an entire text following from
a single inputted prompt. This latter possibility has been the subject of much
public deliberation, especially in response to OpenAI’s development of their
GPT text generators, which effectively operate in such a way (Radford
et al., 2019).

4.3 NLG Systems as Agents

It is here where we shift our consideration of NLG systems from being tools
for manifesting human vision to viewing them as social agents in themselves
(Henrickson, 2018b). Prior to this part of the discussion, it is useful to reflect
upon an earlier argument about agency applied to another text-based
technology: the printing press. The title of Elizabeth Eisenstein’s germinal
work about the impact of the printing press is The Printing Press as an Agent
of Change (1979). Throughout her book, Eisenstein focuses on how the
evolution of print coincided with and catalysed widespread cultural changes
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

related to religion, science, and social order. Eisenstein’s use of the word
‘agent’ instead of ‘tool’ nods to the power of the social structures associated
with the printing press over any one individual’s use of the technology.
Even Gutenberg himself is hardly mentioned. Indeed, the printing press has
been not just a tool, but an institution that has transformed social circum-
stances through new labour economies and processes of text production,
dissemination, and reception. Focusing on a technology’s large-scale impact
rather than singular applications draws attention to that technology’s more
general social and hermeneutic implications. A printed text does more than
just transmit words on a page; it represents an established system of labour
and perpetuates the symbolic value of print that constitutes our current
Order of the Book. The Order of the Book assigns special significance to
books (especially in the printed codex form) because the texts held within
those books direct the operation of social institutions (van der Weel, 2011a:
91). It is therefore essential not just to study printed texts as isolated literary
artefacts, but to recognise the sociocultural development of these texts, as
well as the social and technological developments that have influenced their
forms.
 Reading Computer-Generated Texts 49

So too with computer-generated texts. To appreciate the extent of

algorithmic authorship’s social and literary impact, it helps to recognise
NLG as an agent of change, an institution that transcends any one person or
application. There are, of course, some who take issue with Eisenstein’s
emphasis on a technology as an ‘agent’ in her discussion of the printing
press. One rebuttal by Adriaan van der Weel (2011a: 165) is that ‘however
important the role of the printing press has been in cultural history, however
many genii it may have helped to release from their bottles, it has never –
pace Elizabeth Eisenstein – been more than a passive instrument in the hands
of man.’ Van der Weel urges consideration of the printing press as a tool, an
augmentation to the human experience through conscious human use.
Indeed, the printing press is itself a passive instrument, a tool for actualising
human vision. It is the press’ output – the printed word – and influence that
transforms the press into an agent of change. Put simply, it is not the
technology itself that constitutes agency, but its lasting contributions to
social transformation. In this way, computational systems may also be
regarded as agents of change; the precise nature of change is debatable,
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

but that computational capacity, output, and influence have led to social
change is clear. Another rebuttal to Eisenstein’s work that explicitly decries
her use of the word ‘agent’ declares that ‘[a]gents are humans with will,
intention and responsibility, while agencies are impersonal: corporations,
machines, the weather, gravity. . . . The alternative is to read agent figura-
tively, as a metaphor, more precisely as a personification wherein a thing or
abstraction is endowed with life’ (Teigen, 1987: 8). In either sense, to deem
an NLG system an agent, rather than just a tool, is to acknowledge the
wider-ranging social structures associated with algorithmic authorship
rather than just NLG’s impact on individuals. It is to acknowledge the
transformative social power of computer-generated texts. Thus, ‘agent’ in
this Element refers neither to a human with will nor a thing endowed with
life. It refers instead to participants in always-mediated processes of com-
munication. ‘Agent’ refers to someone – or something – that substantially
contributes to the creation of output that could be construed as meaningful.
‘Agent’ refers to someone or something with influence.
Some NLG systems may be regarded as fitting comfortably within the
lineage of writing tools, as they quite obviously depend upon embodied
 50 Publishing and Book Culture

figures – humans – for their functionality. For other systems, particularly

those wherein the embodied figure is obscured, the distinction between tool
and agent is not so clear. If the Asibot were to autonomously generate its
own text without Giphart’s guidance, could the system still be regarded as
a tool? In a machine learning NLG system, at what point is Margaret
Boden’s assertion of developer responsibility for input determination ren-
dered moot? While some (Bolter, 1991: 189) may argue that in instances of
text generation ‘[t]he computer becomes the author by proxy, in the sense
that it operates on the text as the reader reads,’ a seemingly disembodied
NLG system poses new questions about authorship attribution and how one
may negotiate the hermeneutic contract. This is because authorship attribu-
tion becomes increasingly complex. The widely held notion that ‘computers
only do what they are instructed to do’, proposed by computing pioneer
Ada Lovelace in 1843, still stands (Menabrea, 1843: 722). However, as
cognitive scientist Douglas Hofstadter (1980: 306) notes, ‘you don’t know
in advance the consequences of what you tell a computer to do; therefore its
behaviour can be as baffling and surprising and unpredictable to you as that
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

of a person. You generally know in advance the space in which the output
will fall, but you don’t know details of where it will fall.’ But it is precisely
the unpredictability resulting in original content that prompts a shift in
perceiving some NLG systems from being mere tools for manifesting
human vision to being agents in themselves. Staunch assertions that NLG
systems are no more than tools lack recognition of the potential futures of
development, and hinder discussions of ethical and social circumstances that
may inform judgements of responsibility for system output and claims to
financial gain from that output (Prescott, 2017). Agency is what is per-
mitted. The moment a reader begins searching for the intention driving
a text’s generation is the moment that system is granted agency as a result of
a reader’s attempt to fulfil the conventional hermeneutic contract.
Recognising an NLG system as an agent – and subsequently as an author –
welcomes the NLG system into the realm of what Michel Foucault has
deemed the ‘author function’.
For Foucault, an author’s name is not just a description of textual
ownership and accountability, but also a designation of which texts are to
hold cultural weight. An author’s name allows one to classify and define
 Reading Computer-Generated Texts 51

texts in relation to other texts; those texts comprising an author’s corpus are
seen as complementary, and may be juxtaposed with other corpora.
Foucault’s author function thus recognises the author as a cultural construct
that exists within fluid networks of social circumstances, rather than as any
one individual. In Foucault’s (1998: 221–2) words, ‘the author is not an
indefinite source of significations that fill a work; the author does not
precede the works; . . . [t]he author is therefore the ideological figure by
which one marks the manner in which we fear the proliferation of
meaning.’3 In our world of information abundance, this fear of ‘the pro-
liferation of meaning’ is increasingly present; the responsibility to make
knowledge from information rests increasingly with the reader, who must
swim through a sea of non-linear printed works and digital hypertexts that
are ever more disembodied from singular human sources. To fear the
proliferation of meaning is to fear the responsibility of having to interpret
and negotiate the significance of myriad textual artefacts, in either isolation
or relation to one another. Linking a text with an author – however
imaginary that author may be – streamlines processes of meaning-
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

making, as the reader is not so much faced with a proliferation of meaning

within a text, but an expectation that an individual has written with com-
municative intention. In Foucault’s view, the reader’s assumption of author-
ial identity shrouds the subjective nature of reading, contributing to a sense
that a text has a ‘correct’ interpretation rather than full appreciation of that
text’s contextual malleability.
While writers are physical beings who produce texts, authors are more
abstract cultural constructs. In his elaboration upon Foucault’s article
asking ‘What Is an Author?’, Alexander Nehamas (1986: 686) explains
that writers are ‘actual individuals, firmly located in history, efficient causes
of their texts.’ Authors, however, are ‘postulated to account for a text’s
features and are produced through an interaction between critic and text.
Their nature guides interpretation, and interpretation determines their
nature. This reciprocal relationship can be called, not simply for a lack of

3
Ironically, as some scholars (Wilson, 2004) have observed, Foucault himself
wrestles with the notion of the author as ideological, perhaps due to his own desire
for intellectual recognition and celebrity.
 52 Publishing and Book Culture

a better word, transcendental.’ Despite its transcendentality, the cultural

construct of authorship has very real implications; the common perception
of the author as an individual creative genius informs the meaning derived
from a text. This perception also informs judicial actions related to the
financial and ethical responsibility for a text (Henrickson, 2020). Moreover,
Adriaan van der Weel (2015: 2) has argued for a sociotechnical history of
authorship, writing that ‘the social position of authorship is always founded
in a particular sociotechnical constellation of a particular literate mentality
and a particular dominant technology for the dissemination of texts.’ A text
is always a social artefact, created and received by social participants.
However, those social participants need not be limited to just human
entities. Indeed, social participation may more widely incorporate humans
and computers alike, both of which may be considered social agents.
The notion of computers as social participants is more expressly rooted
in the Computers Are Social Actors (CASA) paradigm, which holds that
humans react similarly to computers as they would to other humans (Nass,
Steuer, and Tauber, 1994; Reeves and Nass, 1996). More recent research
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

into users’ perceptions of Twitterbots affirms CASA’s continued relevance

(Spence et al., 2019). Research in robotics supports the CASA paradigm as
well (Fischer and Bateman, 2006). In CASA’s formative experiments,
researchers observed that participants maintained awareness of interacting
with computers rather than other humans, but nevertheless extended similar
social etiquette (Nass and Moon, 2000). Further, when the experimental
computers ‘spoke’ in different voices, participants responded as through
each voice belonged to a distinct entity; participants also applied gender
stereotypes to these voices. ‘Thus,’ these experiments’ facilitators (Nass,
Steuer, and Tauber, 1994: 77) concluded, ‘computer users who respond
socially to computers do not feel as though they are interacting with
a programmer, but rather attribute socialness directly to the computer
itself.’ However, while computers could be argued to be social actors in
that they prompt subconscious activation of typical social behaviours,
computational entities are not wholly comparable to human entities.
Although users respond similarly to computers as they would to humans,
computers evidently play very different social roles in these situations,
leading to altogether different social dynamics.
 Reading Computer-Generated Texts 53

Actor-network theory (ANT) is useful for exploring such social

dynamics. In ANT, everything in the social and natural world exists within
fluid relational networks. Actors need not be humans. According to ANT
originator Bruno Latour (2005: 71), ‘the questions to ask about any agent
are simply the following: Does it make a difference in the course of some
other agent’s action or not? Is there some trial that allows someone to detect
this difference?’ In Latour’s view, an actor is determined more by its effects
on other actors than by any intrinsic properties. An actor is determined by
its contribution to current contexts. As a result, an actor can never be
considered in isolation, as it acts in accordance with the overarching script,
the cast alongside, and the audience observing the scene. Its utterances, its
realisations of potentialities, must be appropriate to corresponding situa-
tions and recipients. While not necessarily spoken, each of these utterances
has a resonant effect, modifying the current state of affairs through some
sort of detectable difference achieved by interaction with other social
entities. Actors have power because other actors have recognised their
transformative capacities. Actors create unique environments by virtue of
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

their very existence.

Latour uses the words ‘actor’ and ‘agent’ interchangeably, but for the
purposes of this discussion it is worth considering these words’ differing
connotations. An actor performs rehearsed scripts, while agency suggests
accountability and intention. Analyses of computational agency have
tended to use human agency as a baseline for evaluating the perceived
accountability and intention of machines. Some scholars (Johnson, 2006)
have suggested alternative phrasing to accommodate the differences
between human and computer decision-making processes. Others (Stahl,
2004) have argued for comparative approaches to establish a human–com-
puter dichotomy. Others (Powers, 2013) have argued that computers may
one day achieve moral agency by demonstrating phenomenological aware-
ness and intentionality. For all of these scholars, agency is assumed to
correspond with the human characteristic of intention that is (arguably)
informed by free will.
In John Searle’s ‘Chinese Room’ thought experiment, an individual with
no understanding of the Chinese language is locked in a room and given
numerous batches of Chinese writing. The individual is then given a set of
 54 Publishing and Book Culture

English-language rules (a ‘program’) that correlates pieces of Chinese input

with appropriate Chinese responses. Working in this way, the individual
produces output indistinguishable from that of native Chinese speakers.
Computer programs may similarly manipulate formal symbols without
understanding them to produce understandable texts, despite the programs
themselves being incapable of free-will intentionality. Nevertheless, in
Searle’s (1980: 419) words, ‘we often attribute “understanding” and other
cognitive predicates by metaphor and analogy[.] . . . The reason we make
these attributions is quite interesting, and it has to do with the fact that in
artifacts we extend our own intentionality; our tools are extensions of our
purposes.’ As some scholars of algorithmic ethics (Mittelstadt et al., 2016:
10) have observed, though, ‘[w]hen a technology fails, blame and sanctions
must be apportioned. One or more of the technology’s designer (or devel-
oper), manufacturer or user are typically held accountable. . . . Blame can
only be justifiably attributed when the actor has some degree of control . . .
and intentionality in carrying out the action.’ Focus on text production often
centres on intentionality, and Searle himself implies that computers cannot
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

be anything more than tools because they are incapable of the intentionality
demonstrated by humans. Given that natural languages are humans’ pri-
mary means for interpersonal communication, it is reasonable to assume
that those who communicate through natural language are doing so to fulfil
some communicative goal. Searle’s ‘human program’, producing Chinese
texts without understanding the language, may not have so pointed
a communicative goal, and may therefore be considered to lack the assumed
free-will intentionality of conventional authorial agents.
Expectations of human-specific free-will intentionality have also
informed policy discussions related to authorial credit. In the United
States’ 1979 Final Report of the National Commission on New Technology
Uses of Copyrighted Works (CONTU Final Report), Commissioner John
Hersey declares that ‘a definite danger to the quality of life must come with
a blurring and merging of human and mechanical communication’
(National Commission on New Technological Uses of Copyrighted
Works, 1979: 36–7). He argues that one must not equate human beings
and machines because:
 Reading Computer-Generated Texts 55

machines may be able in some distant future to linguistically

‘understand’ but will never be able to experience, never be able
to bring to life, never be able therefore to communicate. Those
aspects include courage, love, integrity, trust, the touch of flesh,
the fire of intuition, the yearning and aspirations of what poets
so vaguely but so persistently call the soul – the bundle of
qualities we think of as being embraced by the word humanity.

In Hersey’s view, empathy – rather than linguistic capacity – distinguishes

human writers from NLG systems. Empathy (and, in turn, the ‘soul’)
establishes textual value, appealing to human readers who too exhibit
empathy. Extending this argument, the CONTU Final Report stresses
that the computer is incapable of consciously intending communicative
meaning; computers are considered tools for manifesting human vision,
and do not wield their own intention. The Report (National Commission on
New Technological Uses of Copyrighted Works, 1979: 45) summarises this
argument thus: ‘Computers are enormously complex and powerful instru-
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

ments which vastly extend human powers to calculate, select, rearrange,

display, design and do other things involved in the creation of works.
However, it is a human power they extend.’ The human power extended
is implied to be one of free-will intentionality.
An argument for computational agency questions this focus on the
assumed free-will intentionality of authorial agents by following from
Daniel Dennett’s notion of the ‘intentional stance’. Through the intentional
stance, the behaviour of an actor - and by extension an agent - is interpreted
by considering its beliefs and desires (Dennett, 1971). When faced with an
entity – whether human or not – we assume rational behaviour appropriate
to the relevant social circumstances. We therefore ascribe beliefs and desires
to that entity. Dennett uses a chess-playing computer as an illustration.
‘Lingering doubts about whether the chess-playing computer really has
beliefs and desires are misplaced,’ he writes (Dennett, 1971: 91), ‘and
whether one calls what one ascribes to the computer beliefs or belief-
analogues or information complexes or Intentional whatnots makes no
difference to the nature of the calculation one makes on the basis of the
ascriptions.’ Nevertheless:
 56 Publishing and Book Culture

We do quite successfully treat these computers as

Intentional systems, and we do this independently of any
considerations about what substance they are composed of,
their origin, their position or lack of position in the com-
munity of moral agents, their consciousness or self-
consciousness, or the determinacy or indeterminacy of
their operations.

For Dennett, users’ perceptions of systems are more significant than actual
system functionality. Following from ANT, if users ascribe beliefs and
desires to a system, thereby welcoming the system into a social network,
that system becomes a social agent. In one example of the intentional stance,
Dennett considers a thermostat. One may regard a thermostat as a mere
control mechanism for regulating temperature, whose regulatory function-
ality may be applied to myriad situations (e.g. a car’s cruise control). The
thermostat’s technology is not constrained to any one boiler, or even to the
domain of temperature. However, the more the thermostat is integrated into
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

the world – the more complex its internal operations become as a result of
enriched connections with other entities associated with its assigned task
(e.g. the thermostat purchases its own boiler fuel or checks a house’s
weather stripping) – the more one can say that the thermostat ‘has beliefs
about heat and about this very room, and so forth, not only because of the
system’s actual location in, and operations on, the world, but because we
cannot imagine another niche in which it could be placed where it would
work’ (Dennett, 1987: 31). The thermostat becomes a distinctive agent that
fulfils a certain social function, prompting the emergence of new organisa-
tions of behaviour and thought that centre on the thermostat. David
Herman’s (2008: 257) more recent consideration of the intentional stance
applies it to narrative theory, concluding that ‘stories simultaneously reflect
and activate a disposition to adopt the intentional stance.’ Connecting
Dennett’s notion of the intentional stance with more recent developments
in AI and algorithmic authorship, there are efforts in the field of explainable
AI/explainable computational intelligence to enable machine learning sys-
tems to describe their decision-making processes, often in natural lan-
guages, and these descriptions may support a sense of autonomous
 Reading Computer-Generated Texts 57

computational intentionality. Such systems stimulate reorganisations of

behaviour and thought motivated by efforts to understand computational
decisions. An NLG system may not just reflect human beliefs, but may be
regarded as having beliefs of its own because it becomes a distinctive entity
that fulfils a certain social function resulting in original output. These beliefs
are not necessarily rooted in free-will intentionality. Many systems exercise
forms of intentionality not governed by free will, but by efforts to fulfil
designated objectives, like Searle’s human program. Systems become social
agents by contributing to and transforming social networks.
Such scrutiny of semantics may seem pedantic, but in a discussion of
NLG systems the recognition of computational agency is imperative for
complete understanding of these systems’ roles within contemporary cir-
cumstances. Regardless of whether an author is a human or a computer, the
reader’s process of meaning-making is fundamentally the same: fashion
a narrative about authorial intention. As Katherine Hayles notes in her book
about Digital Subjects and Literary Texts (2005: 197) when discussing the
attribution of motives and intentions related to texts, ‘narratives allow us to
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

construct models of how others may be feeling and acting, models that
coevolve with our ongoing interior monologues describing and interpreting
to ourselves our own feelings and behaviors.’ Writing about computer-
generated texts in particular, Manuel Portela (2018: 195) notes that ‘[e]ven if
there is a certain degree of mathematical randomness in the verbal output,
linguistic combinations will have emergent meanings that will be read
literally. Random textual instantiations thus open up machinic constraints
to the unconscious of the reader.’ All this is to say that the narrative text – in
this case a computer-generated text – prompts the production of another
narrative about that text within the readerly mind. The reader seeks
authorial intention through engagement with engrained hermeneutic pro-
cesses, fulfilling Dennett’s intentional stance. If a reader who did under-
stand Chinese were to read the text produced by Searle’s human program,
for example, that reader would still consider the text in light of the assumed
intentionality of an imagined authorial agent. We legitimise our lived
experiences to ourselves and others by using natural language, and the
use of natural language alone is enough to provoke instinctual ascribing of
beliefs and desires to an imagined author so as to discern the assumed
 58 Publishing and Book Culture

communicative function of a text. The NLG system’s use of natural

language therefore immediately distinguishes it as being on the figurative
border between tool and agent. The computer may be used to alter the
current state of affairs through an ‘actor’ role, as in Whalen’s Several Houses
NaNoGenMo submission, wherein the computer was employed as a tool.
Much algorithmically authored output, however, situates the computer
within the realm of ‘agent’ as a result of the social power assigned to it,
however unconsciously, once a reader attempts to discern the communica-
tive function of output by imagining an author. This is a question of social,
rather than moral, agency.
It is this permitted agency that links Dennett’s intentional stance to the
hermeneutic contract. Recall that the hermeneutic contract refers to the
relationship between the reader and the author. It rests on two assumptions:
that readers believe that authors want them to be interested in their texts,
and that authors want readers to understand their texts. To fulfil this
contract, however, the reader must assign beliefs and desires to the dis-
embodied author. As discussed in Section 3, the modern notion of ‘author’
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

tends towards understanding the author as an individual creative genius,

imagined as per a ‘cult of personality’. Yet, to cite NLG system developer
Jean-Pierre Balpe (1995: 24 [translation my own]), if ‘computer literature is
literally readable, then it is radically positioned outside of this “classical”
literary ideology wherein a master of letters has reigned for centuries: no
inspiration or original experience, no intention or genius . . . no authorial
individuality.’ Regardless of whether the author is a human or non-human
entity, the reader infers the intentions of an imagined and disembodied
author based on a process of meaning-making that rests upon the assump-
tion of mutual understanding through shared sociocultural traditions. To
investigate intention by necessity assumes agency. It is therefore not
enough to merely say that NLG systems are tools, or even social actors.
They are social agents, because readers treat them as such.

4.4 Algorithmic Authorship and the Hermeneutic Contract

Readers are rarely given detailed information about processes of computer-
generated text production. However, the results of a recent series of studies
exploring how authorship is attributed to computer-generated texts indicate
 Reading Computer-Generated Texts 59

that readers often identify NLG systems as authors (Henrickson, 2019a/b).

After all, when one reads the final Ik, robot story co-produced by Ronald
Giphart and the Asibot, one cannot distinguish which parts of the texts have
come from Giphart and which have come from the system. The collabora-
tion between human and computer has resulted in a seamless ‘synthetic
literature’ that reflects both Giphart’s and the Asibot’s lived experiences and
intentions, as well as their relationships with one another. There are two,
rather than one, sources of communicative intention: Giphart, in corporal
form; and the Asibot, in computational form. Giphart’s intentions may
reflect the free will that is presumed of human individuals. The Asibot’s
intentions, though, may constitute a wholly different kind of intention: one
characterised by efforts to fulfil designated objectives through processing
and repurposing Dutch-language e-books for the production of sufficiently
original content. The distinction between these two forms of intention
permits more nuanced consideration of the ways in which algorithmic
authorship may contribute to textual landscapes. The Asibot’s – and,
more generally, the algorithmic author’s – power is exerted throughout
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

the co-creative process and is affirmed the moment a reader begins inter-
preting its output. The writing system, human writer, and readers exist
within a fluid network of social relationships rooted in reciprocal exchanges
related to power, whether explicitly or implicitly acknowledged. This is not
to neglect the important questions of moral, financial, and legal responsi-
bility for textual content. Recognising an NLG system as a social agent does
not automatically make that system solely responsible for its output; ques-
tions of responsibility are still to be sufficiently pondered. Recognising an
NLG system as a social agent, though, does allow for the situation of NLG
within a lineage of textual technologies and literary studies. Readers con-
sider computer-generated output in light of their expectations for human-
written texts. ‘When faced with a totally new situation,’ Marshall McLuhan
(with Fiore and Agel, 1967: 74) writes, ‘we tend always to attach ourselves
to the objects, to the flavor of the most recent past. We look at the present
through a rear-view mirror. We march backwards into the future.’
Computer-generated texts bring conventional understandings of the
author–reader relationship – what I have called the hermeneutic contract –
into question. The author of a computer-generated text is ambiguous:
 60 Publishing and Book Culture

potentially human, potentially computational. Nevertheless, the understanding

of texts as being of social – interpersonal, human – value persists. Computer-
generated texts not only appeal to readers’ expectations of textual authority
simply by virtue of being natural language text, but also by mimicking the
familiar forms of the printed word. Despite computer-generated and human-
written texts often being indistinguishable, though, the two means of text
production are decidedly different. This is why it is not enough to consider
the implications of NLG systems based on output alone. In McLuhan’s words:

All media work us over completely. They are so pervasive

in their personal, political, economic, aesthetic, psychologi-
cal, moral, ethical, and social consequences that they leave
no part of us untouched, unaffected, unaltered. The medium
is the massage. Any understanding of social and cultural
change is impossible without a knowledge of the way media
work as environments. All media are extensions of some
human faculty – psychic or physical (McLuhan, Fiore, and
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Agel, 1967: 26).

To appreciate the extent of algorithmic authorship’s sociocultural impact,

one must consider the unique labour economies and social infrastructures
spurred by NLG. Although recent discussions about computer-generated
texts have focused primarily on whether computer-generated texts can pass
the Turing test – whether a human reader can be tricked into believing that
a computer-generated text has been written by a human – such a measure
only evaluates final linguistic output and does not appraise computational
thought, intelligence, or intentionality. New technologies like NLG prompt
renegotiation of the self and the self in relation to society, encouraging
reassessment of human roles in established processes of production and
reception. Arguably, the Turing test standard detracts from the more
important issue of determining algorithmic authorship’s place within the
current sociocultural and literary climates. It detracts from attention to the
human experiences of computer-generated texts.
It is through language that interpersonal agency is exercised, and the
hermeneutic contract is rooted in the reader’s expectation of authorial
 Reading Computer-Generated Texts 61

agency that is informed by lived experience. Texts are perceived as products

of human self-expression. Of course, genre may also inform one’s expecta-
tions of textual content. This is why, for example, Zach Whalen’s The
Several Houses of Brian, Spencer, Liam, Victoria, Brayden, Vincent, and Alex
is recognisable as a piece of children’s literature, despite its occasional adult
themes and jumbled illustrations. The reader reads according to the con-
straints of the customary. Still, though, it is the reader who establishes and
maintains the hermeneutic contract through acts of interpretation, and the
hermeneutic contract stands regardless of genre; it simply holds that texts
are understood as having been produced from communicational intention.
What is more, interpretation, in Alexander Nehamas’ (1986: 688) words,
‘places a text within a perpetually broadening context, not within
a continually deepening one. . . . Interpretation ends when interest wanes,
not when certainty, or an ultimate meaning, is reached.’ A reader may
regard a text as the product of authorial intention, but it is the reader who
actually determines what a text means.
So, how important is recognition of authorial intention for textual
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

interpretation? This question has long been debated by literary critics,

who have cycled between tendencies towards authorial intention (e.g.
biographical criticism), and then towards reader experience (e.g. New
Criticism). Writing from a perspective informed by both book history
and textual scholarship, Peter Shillingsburg (2006: 55–6) explores the
ambiguous nature of intention. ‘[N]o matter how radical one’s objection
is to the concept of intention as a controlling element of textual interpreta-
tion, some aspect of intention by some agent of intention is inescapable in
any reading act,’ Shillingsburg writes. ‘[R]eaders knowingly or blindly
participate in a negotiation over authority in script acts, finding or inventing
concepts of authorship and intention, which they then ignore or flout or
seek to embrace.’ The hermeneutic contract recognises the complex nego-
tiations that occur within the reader to discern authorial intention while still
establishing one’s own interpretation of the text under consideration. There
is, of course, the reader’s expectation of the text as something intended to be
understandable and interesting or, in some instances, intentionally incom-
prehensible or dull for aesthetic purposes. Texts are not simply aesthetic
artefacts, though. They are embedded with communicative function and, as
 62 Publishing and Book Culture

Shillingsburg rightly observes, the precise ways readers may negotiate this
communicative function are not so easily anticipated. One oft-referenced
stance is the ‘death of the author’: Roland Barthes’ 1967 argument that
attributing an author to a text constrains its readers’ interpretations of that
text given an assumption that there is one definitive meaning established by
the writer. Barthes (1977: 148) explains that ‘[c]lassic criticism has never
paid any attention to the reader; for it, the writer is the only person in
literature. . . . [W]e know that to give writing its future, it is necessary to
overthrow the myth: the birth of the reader must be at the cost of the death
of the Author.’ But although the printed word is an artefact largely
disembodied from its initial writer through the mechanised and communal
processes of its production, the sense of authorial intention is an integral
part of its legitimacy. As a reader reads, she interprets a text in light of an
assumed communicational intention informed by a commitment to mutual
understanding that itself is informed by shared sociocultural traditions and
appropriate language usage. Whether the reader chooses to accept what she
believes was the author’s intention during writing, or whether she chooses
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

to disregard it, meaning is discerned through attention to intention. To

reiterate Shillingsburg, the reader consciously or unconsciously negotiates
a text’s authority, choosing to succumb to or ignore perceived authorial
intention. The author is not dead, but imagined. And the author is impor-
tant, for it is in relation to the author that the reader makes meaning. It
matters little whether the author is human or algorithmic; the reader
continues to search for meaning.
Section 3 argued that algorithmic authorship extends the current societal
trend towards what I have called ‘hyper-individualism’. Hyper-
individualism refers to each individual’s tendency towards self-
gratification, driven by senses of self-importance and uniqueness.
Through individualism, each seeks personalisation through subjective cura-
tion of media; through hyper-individualism, personalisation is pervasive in
both analogue and digital forms. While media may still be curated by
individuals, curation is now subject to algorithmic filters that are both
conspicuous and concealed. A comprehensive analysis of computer-
generated texts must therefore account not only for the author function,
but also the unique functions of the reader and the text itself in light of the
 Reading Computer-Generated Texts 63

proliferation of non-experts: ordinary people reading and interpreting texts,

who may have limited knowledge of these texts’ production processes.
Hyper-individualism not only legitimises an individual’s sense of self, but
also supports fragmented processes of meaning-making wherein the inter-
pretive burden rests ever more with the reader rather than with the author as
a result of algorithmic ubiquity.
Hyper-individualism is magnified by the rise of companies that specialise
in the generation of news articles for consumption by general and niche
audiences alike. When transforming datasets (e.g. sports scores, weather
forecasts, stock performances) into readable narratives, these companies’
systems demonstrate the capacity to rapidly generate texts that are highly
personalised in both content and register. Computer-generated texts may
therefore appear disembodied from their producers, but find embodiment
within the readers for whom they were generated. Readers not only
consume relevant knowledge, but help create it. Even for non-
personalised computer-generated texts the individualism of the reader is
affirmed, if only because there is no longer an obvious author to whom one
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

may attribute authorship and, by extension, intention. The author function

blurs.
Attention to readers’ documented experiences of computer-generated
texts is especially important given that some NLG systems have begun
producing altogether new kinds of literature: literature that may not con-
form to the syntactic and semantic conventions to which readers have
grown accustomed. Considering one of the short stories in Racter’s 1984
The Policeman’s Beard is Half Constructed, marketed as ‘the first book ever
written by a computer’, Scientific American writer A. K. Dewdney (1985: 10)
writes:

Even a layperson may have concluded that Racter is decid-

edly schizophrenic. On the one hand, the ideas brought
together in these sentences seem to form a coherent whole.
It is nice that Sarah and Bill sing to each other. Although
I had not thought of singing as dangerous, I am positively
charmed by the idea of the two eating lamb, stroking each
other and chanting about their ups and downs. On the other
 64 Publishing and Book Culture

hand, the allowances I have been making for Racter all along
are stretched to the breaking point when Racter mentions
that besides their love they also have typewriters. Invited to
share in this extraordinary insight, I tremble on the brink of
a completely unknown mental world, one that I would
prefer not to enter.

Racter’s syntactic bumbling, jumping from one topic to another, often with
no logical connections between topics, leaves Dewdney with the sense that
he is reading the work of an author unhinged. Nevertheless, Dewdney
acknowledges the ‘extraordinary insight’ that this output might offer him,
despite his reluctance to enter into the computational writer’s mind.
In another review of computer-generated text, Charles Hartman (1996: 71)
recounts his experiences developing a poetry generation system, explaining
why the output of his AutoPoet could not be deemed a ‘victory’. For context,
the AutoPoet generated such output as:
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

The garden of steel – place – had figured in

this. When I am every afternoon,
how can’t the last teacher write? But I
was art without my play between a result
and the metabolism, and the night
of language toward a story between the part
and any light (the thin subject) remains.

And so forth. Hartman (1996: 72) reflects:

AutoPoet embodied an inappropriate idea of poetry. As

long as the goal was the imitation of a human poet – or as
long as the poem’s reader was encouraged to think that was
the goal – I wasn’t likely to get any farther. What’s wrong
with the AutoPoetry I’ve quoted here (and all the other
reams of it the machine would produce until it was turned
off) is exactly that it’s imitation poetry. All our habits of
reading are called upon, all the old expectations, and then let
 Reading Computer-Generated Texts 65

down. ‘Monologues of Soul and Body’ [another one of

Hartman’s computer-generated poems] had worked because
its ‘body’ sections were so different from human poetry. It
had successfully demanded its own way of reading.

‘All our habits of reading are called upon, all the old expectations, and
then let down,’ Hartman declares. However, this Element, supported by
a series of empirical studies investigating reader responses to computer-
generated texts (Henrickson, 2019a/b), shows that these expectations are
not let down. Readers continue to engage in processes of meaning-
making when faced with computer-generated texts, however dull or
nonsensical those texts may be. Readers continue to imagine the author,
even for a computer-generated text, because they instinctually aim to
fulfil the hermeneutic contract to which they are accustomed. Although
Hartman believes his AutoPoet failed because it did not sufficiently draw
from its own lived experiences and emotional perceptions to produce
truly ‘unique’ work, the only expectations that are actually let down are
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

those of Hartman himself.

All the same, the expectation for computer-generated texts that pass the
Turing test is pervasive amongst those considering algorithmic authorship.
Underlying such an expectation is an assumption that the end goal of an
NLG system is not to generate output that is unique in its own right, but
output that can deceive human readers into believing that it has been
human-written. Martin Eve (2017: 46) reflects upon an RNN he trained to
generate textual output that mimics the writing style of articles from the
literary studies journal Textual Practice: ‘the network has no motivation
towards communication and no epistemological goal except to achieve ever
more perfection in its stylistic mimicry of the articles in Textual Practice.’
Despite the RNN producing text that is understandable, albeit verbose, Eve
is unwilling to call the system an ‘author’. He calls the proliferation of NLG
systems ‘alarming’, not only because these systems threaten to wholly
mechanise current labour practices, but also because he believes that they
depreciate narrative quality and undermine the work of narrative scholars.
According to Eve (2017: 51):
 66 Publishing and Book Culture

[I]n achieving a mimesis of human writing – remember,

a measure of intelligence formed only by anthropocentric
reference to the human – computational writing asks us to
imagine a world in which there are no more humans under-
taking such labour. Such thinking only emerges, though, in
the imagined substitution of the human with human-like
automata. This imagined world is both a post-
anthropocentric world for writers and a world in which
a writing machine that is legitimated by human-like char-
acteristics is inscribed at the centre. It is concurrently
a world in which we have no benchmark of contemporary
writing success, but one that is nonetheless dominated by
machines that meet that nostalgic target.

Eve’s hypothetical world exemplifies an approach to thinking about com-

puter-generated texts that tends towards apocalyptic outcomes for human
labour because computers can produce adequate imitations of human-
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

written texts. Reviewing Racter’s The Policeman’s Beard nearly two decades
after its publication, experimental writer Christian Bök (2002: 10) more
bluntly writes that ‘RACTER is a mindless identity, whose very acephalia
[absence of a head] demonstrates the fundamental irrelevance of the writing
subject in the manufacture of the written product. The involvement of an
author in the production of literature has henceforth become discretionary.’
Similarly, novelist and critic Steve Tomasula (2018: 50) has argued that
computer-generated texts exemplify a kind of ‘postliterary literature’ that ‘is
informed by an accompanying posthuman ethos – one that is at odds with
an ethos based upon the uniqueness of the individual, and its cousins,
especially originality.’ But an algorithmic author does not ask us to imagine
a world in which there are no more humans undertaking textual labour. It
simply prompts us to imagine a world in which algorithmic authorship has
a place alongside human authorship. It encourages us to adapt to the new
circumstances of an increasingly digital age. And, most importantly, NLG
systems may produce different kinds of texts altogether: ones that challenge
readers’ syntactic and semantic expectations and demand their own inter-
pretive approaches. As a result, value judgements about system output are
 Reading Computer-Generated Texts 67

impractical at this time. As new genres emerge, it is more appropriate

simply to observe creators – both corporal and computational – as they
explore the new social and literary potentialities permitted by algorithmic
authorship. This Element has aimed to start a discussion about how modern
conceptions of authorship and reader responsibility apply to instances of
computer-generated texts. Once this foundation has been established, jud-
gements of literary value – as well as the perceived extent and value of
authorial intention – may be of greater relevance.

4.5 Conclusion
Disembodied from the text, save for perhaps an accompanying name, the
author has long been imagined by the reader through participation in the
hermeneutic contract. For computer-generated texts, though, the author is
not so clear a figure, not so easily personified through consideration of
linguistic tone or perceived interpersonal connection. Yet the reader’s
assumption of the conventional hermeneutic contract continues to inform
the reading experience, and the NLG system is evaluated through compar-
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

ison to the human writer. When a system produces output that exists
alongside human-written text, readers may interpret this output as reflecting
the same commitment to adhere to and perpetuate shared sociocultural
traditions as that presumed of a human writer: in such instances, these are
the traditions of appropriate semantic and syntactic structures. In this way,
the NLG system itself takes on a fundamentally social role distinct from that
of its developers. It stands on the fence between tool and agent, teetering
ever closer to the realm of agency that characterises modern conceptions of
authorship. Recognising NLG systems as agents in themselves, as social
institutions with transformative power rather than mere tools for actualising
human vision, serves to counteract comparisons of NLG systems to human
writers, permitting clearer recognition of the unique contributions of
algorithmic authorship to the social and literary landscapes. The type of
agency exerted by these systems need not be characterised by the free-will
intention of human writers, but may instead be characterised by the
programmed intention to fulfil a designated objective that is respected
within wider social networks: a deterministic approach to being, so to
speak. Algorithmic authorship necessitates new infrastructures of text
 68 Publishing and Book Culture

production and dissemination, but also necessitates new infrastructures of

thought. Computer-generated texts necessitate a new literacy that helps us
make sense of digital methods of text production and articulate our own
human experiences and interpretations using digital media. This is a literacy
that more explicitly recognises the reader’s responsibility for making mean-
ing. The hermeneutic contract maintains its relevance, but only as a starting
point for fuller understanding of algorithmic authorship’s implications
achieved through a sociology of computer-generated texts.
A call to recognise NLG systems as agents in themselves does not negate
the fundamentally human facets of text production and interpretation.
Computer-generated texts – like human-written texts – always emerge
from, and function within the confines of, human vision. Systems are
developed according to developers’ objectives, and texts are read with
expectations of underlying communicational intention. Texts have no
inherent meaning, but are assigned meaning by human writers and readers.
Further, computational agency complements human agency, facilitating
new methods of exploring and articulating human experience. NLG systems
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

augment the speed and power of human capacities for text production,
extending human faculties. While they may take the place of some human
writers, they remain dependent upon humans for their functionality, opera-
tion, and maintenance. They contribute to new sociologies of text, offering
alternative textual performances wherein the reader’s experience is priori-
tised over that of the author. Acknowledging NLG systems as agents rather
than mere tools prompts reflection upon the distinct labour economies
driving and emerging from algorithmic authorship. Such acknowledgement
facilitates recognition of the transformative social power of system
output, contributing to more nuanced discussions of how algorithmic
authorship conforms to and/or confronts conventional author–reader
relationships.
And so we return to Amersfoort, in de Bibliotheek Eemland. A robot
arm lifts a copy of the new edition of Ik, robot, passing it to a man with an
advanced prosthetic hand: a cyborg. The cyborg then passes the book to the
human, who raises it for the audience to see. The crowd cheers, excited to
read the synthetic literature that is the product of human–computer colla-
boration. NLG systems such as those used to write a new story for Ik, robot
 Reading Computer-Generated Texts 69

may be equated with puppets, generating output according to specific

instructions, but their implications for authorship make them something
more. These systems represent a new literary institution wherein the human
author is overcome by a computational entity, and the reader is at the
forefront of meaning-making. Of course, reading is never a passive act, and
it never has been. In instances of computer-generated texts, though, the
reader builds upon computational output to (co-)create knowledge from
information, to embed texts with meaning and social relevance. Through
human–computer synthesis, a text is generated according to appropriate
literary structures. By employing language from which readers derive
meaning, the NLG system becomes a social agent in itself. It is an agent
of change.
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press
 70 Publishing and Book Culture

5 Conclusion
Let us now revisit Simon, sitting in his local café with his newspaper set
down on the table in front of him. ‘Majority of New Mothers in
Wolverhampton Are Unmarried’ (Anon., 2017) reads the title of one
story. Rather than being written by a human, though, this story has been
computer-generated by a company called Urbs Media, which boasts its
technological power for ‘crafting stories and harnessing automation to mass
localise’ (Urbs Media). Now I feel like I’m on the set of a sci-fi film, Simon
thinks to himself.
Despite sounding science-fictional, computer-generated texts abound.
Data-driven news articles like Simon’s ‘Majority of New Mothers’ are
common applications of NLG. However, as has been shown throughout
this Element, news is only one of NLG’s many domains. Machines are
generating texts for expository and aesthetic purposes, and have been doing
so for more than a century.
Computer-generated texts in their current forms, however, bring mod-
ern conventions into question. In our current state – the Order of the
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Book – books (especially tangible codices) are assigned symbolic value

because the texts contained within them direct the operation of social
institutions that sustain modern representative democracy and perpetuate
cultural and literary heritage (van der Weel, 2011a: 91). The Order of the
Book is characterised by social institutions of authorship and widely held
understandings of reader responsibility for interpreting messages that
authors have explicitly and implicitly embedded within their texts. Yet as
van der Weel (2011a: 178) observes, the ‘interpretive burden’ rests increas-
ingly with the reader, as digital-born texts challenge the standardisation and
linearity of the mass-printed word. Left to make sense of an abundance of
available texts, the reader is put in a state of privilege that emerges from and
supports a cultural tendency towards hyper-individualism: an exaggerated
understanding of the reader as meaning-maker. NLG augments hyper-
individualism through its ability to mass-generate highly localised texts in
short periods of time. One example of a highly localised text is Simon’s
‘Majority of New Mothers’ article, which reports statistics specific to
Wolverhampton. In more extreme cases, individual readers are presented
 Reading Computer-Generated Texts 71

with narratives produced just for them, affirming their senses of self within
an age of information overload that may otherwise seem stifling to any one
voice. Although the Order of the Book is challenged by the sheer prolifera-
tion of texts that permit hyper-individualised reading, NLG developers
capitalise upon ingrained assumptions of textual and codicological value.
Computer-generated texts represent today’s digital ecology, transform-
ing what has largely been seen as an extension of the human self – text, the
claim to authorship – into an esoteric entanglement of human and computer
involvements. Ultimately, NLG is about computational choice and creativ-
ity. According to its programmed instructions, the system must choose what
content to include in its output, and how to convey selected output. By
applying a sociological perspective to an analysis of computer-generated
texts, though, we can see these texts more clearly for what they are: human
artefacts, produced by humans for human reception. These artefacts have
notable implications for the social and literary spheres, with one of the most
significant but implicit being that they bring into question the conventional
imagined relationship – the hermeneutic contract – between reader and
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

author. I have therefore argued for a semantic shift from the NLG system as
tool to the system as agent, based on a broader understanding of agency as
defined by both free-will and programmed forms of intentionality. The
NLG system holds social power because readers attempt to derive meaning
from its output. The NLG system is a communicative medium with the
ability to transform social circumstances with its output, and even its very
presence. It has social agency; it is a social agent; it truly becomes an
algorithmic author.
In conversation at the Electronic Literature Organization’s 2018 con-
ference, one artist whose practice centres on producing computer-generated
text solicited advice.4 ‘When my work does get press,’ she explained, ‘the
output is often attributed to my bots, rather than to me. But I am the one
who creates. How do I make sure I am the one being given credit?’ This
leads to what are possibly the most important questions arising from this
discussion, which have gone unaddressed until this point. What are we to
do with this theorising? How do we want readers to respond to these

4
Anonymised poet, conversation with the author (15 August 2018, Montreal).
 72 Publishing and Book Culture

works? Can developers of NLG systems guide readers towards desired

responses? How do different genres of computer-generated texts – news
articles and poems, for example – demand different kinds of interpretive
responses or authorship attributions? What does accepting an NLG system
as a social agent mean for literary analysis of both expository and aesthetic
texts? It is beyond the scope of this Element to provide such answers, and
the responsibility lies with future research projects involving both scholars
and system developers to settle on answers that are rooted in nuanced
understandings of where computer-generated texts fit within current cir-
cumstances. As computer-generated texts become increasingly present in
our everyday lives, these are questions that demand urgent consideration.
The answers to such questions will inform the bot-making artist’s practice
and journalistic approaches to authorship attribution.
Indeed, studies about the social implications of algorithmic authorship
are nearly non-existent. Discussions about NLG have been dominated by
computer scientists driving system development, and often focus on eval-
uating the functionality of systems using computational and empirical
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

methods; functionality is more often than not defined by a system’s ability

to deceive users into believing its output has been human-written. The
preceding pages represent an effort to begin filling this gap by discerning
the social implications of algorithmic authorship and establishing
a theoretical foundation for further studies. This effort was channelled by
situating NLG’s lineage and current applications within greater historical
contexts related to the effects of technological development on communica-
tions media. Despite uncertainty about this technology, NLG is only
becoming more prominent in the modern textual landscape, which itself is
ever-changing and subject to cartographic subjectivity.
The social, hermeneutic, and literary implications of algorithmic author-
ship are vast, and research incorporating scholars from the humanities, the
social sciences, computer science, and other academic fields and industrial
areas is needed to fully comprehend the implications of this textual phe-
nomenon. As Dennis Tenen writes in his book The Poetics of Computation
(2017: 199), ‘[t]echnologies that govern literacy cannot be allowed to
develop apart from the humanities. Such detachment threatens the legacy
of interpretive practice, enacted on page or pacemaker. Whatever their
 Reading Computer-Generated Texts 73

politics, literary scholars, philosophers, and historians must negotiate the

tactics of reading alongside lawmakers and software engineers.’ Tenen
advocates interdisciplinary collaboration to negotiate how technologies
may create, reinforce, and subvert hierarchical systems of social governance
privileging a technocratic elite. This Element has set the groundwork for
such negotiations as they pertain to algorithmic authorship. Starting from
the broad base established herein, future research can approach the world of
computer-generated texts with narrower scopes that explore more pointed
research aims and objectives.
Future studies, for example, could place greater emphasis on the exegetical
experiences provoked by computer-generated texts, more precisely delineat-
ing readers’ processes of making meaning from computational output. Such
studies may incorporate literary analyses of these texts, exploring the stylistic
elements of computer-generated texts that influence conscious and uncon-
scious processes of interpretation. They may draw from the scholarship about
electronic literature – works where the digital not only supports the writing
process, but plays an integral part in a text’s development and reception – as
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

such scholarship has largely focused on the literary merits of specific texts,
with occasional investigations into the phenomenological experiences of
readers engaging with them. Future studies may also benefit by drawing
from the field of affective NLG, which is defined by its efforts to evoke
emotional responses in readers through conscious employment of words and
sentence structures. However, the well of empirical studies related to affective
NLG and the aesthetic experiences prompted by computer-generated output
is remarkably shallow, and has been largely untouched by researchers outside
of computer science. Supplemented by a team that incorporates the electronic
literature community and affective NLG researchers, humanities researchers
are especially well suited to catalyse this area of study, given their attention to
the subtle ways language may be used for persuasion, entertainment, and
aesthetic purposes. By integrating these disciplines’ quantitative and qualita-
tive tendencies, respectively, a deeper understanding of NLG’s potential for
overcoming common perceptions of mechanised impassivity may be realised.
Humanities researchers are also well suited to scrutinising the metaphors
both developers and ordinary readers use to talk about NLG systems. The
humanities have long been attuned to the ways in which metaphor reflects
 74 Publishing and Book Culture

and shapes perceptions of the world, and close attention to perceptions as

they are manifest through language may reveal more tacit understandings of
algorithmic authorship. Metaphor may be helpful in positioning algorithmic
authorship within the current cultural climate, but may also impede full
comprehension of technological capability. Further, different groups –
ordinary readers and developers, for example – may use different meta-
phors aligned with their respective frames of reference. On a practical level,
though, understanding – and capitalising upon – the metaphors ordinary
readers already use to conceptualise algorithmic authorship may contribute
to more effective descriptions of systems that apply user-friendly metaphors
instead of more potentially confusing developer-driven ones. On a more
theoretical level, investigation into the metaphors that ordinary readers use
may promote deeper understanding of attitudes towards algorithmic
authorship, whether they be positive, negative, or (as I predict) altogether
uncertain. Literary scholars are especially well equipped for this task.
Future studies could also shift the focus from the reader – who was
paramount here – to the developer, investigating intention expressed
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

throughout development, as well as distinguishing the underlying assump-

tions that inform a system’s course of development. Such studies may
benefit from formal qualitative and quantitative research, as well as from
less formal observations of development team meetings and interviews with
individual developers. Working with development communities would
undoubtedly reveal entirely different perspectives on the social and herme-
neutic implications of algorithmic authorship: perspectives that may affirm,
contradict, or add nuance to those of ordinary readers, contributing to more
holistic understandings of algorithmic authorship’s place in the modern
world.
Similarly, examination of different language groups’ responses to com-
puter-generated texts may lead to alternative conclusions. The French-
speaking world, for example, has become deeply familiar with experimental
literature driven by systematic production. The Oulipo movement, ongoing
since 1960, has prompted academic and popular discussion about text-
production processes; the published proceedings of a 1994 conference in
Paris (Vuillemin and Lenoble, 1995) show that the French have long been
discussing questions of authorship as they pertain to computer-generated
 Reading Computer-Generated Texts 75

texts, albeit only occasionally. As described in Section 4, the Dutch-

speaking world recently saw the release of a new edition of Isaac
Asimov’s Ik, robot (I, Robot) that featured a new story co-produced by an
NLG system; the sense of excitement at this book’s launch event suggested
generally positive reception of a computer-generated text. The distinct
literary traditions of different language groups may inform these groups’
responses to algorithmic authorship, and may even contribute to deeper
appreciation of computer-generated texts on the part of readers from these
groups. Language idiosyncrasies may also impact speakers’ uses of meta-
phor to conceptualise algorithmic authorship.
There are also avenues for future studies of algorithmic authorship
within the domain of ethics. As NLG systems are developed to be increas-
ingly autonomous, exploration of the potential ethical quandaries posed by
the proposed treatment of NLG systems as social agents – rather than just
tools for manifesting human creative vision – will become paramount. How
is remuneration to be paid to a computer system that (currently) has no
financial interest? If no payment is necessary, how are the economic
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

structures of publishing affected? Who is to be held accountable for

offensive or inaccurate texts when those texts have been generated by
seemingly autonomous systems? A discussion of whether AI systems and
robots ought to be afforded moral agency is already underway (Bryson,
2018). This discussion, however, has hardly been extended to text produc-
tion, and there has been virtually no scholarly or industrial consideration of
the effects of computer-generated texts on the current architectures of the
publishing industry aside from some legal deliberations about copyright law
and intellectual property. Recognising an NLG system as a social agent
does not make that system morally, financially, or legally responsible for its
output; it simply serves as a starting point for recognising the unique
contributions NLG systems can make to text production. Questions about
the authorial responsibility for computer-generated texts have yet to be
sufficiently pondered, despite attempts by such bodies as the United States’
1970s Commission on New Technology Uses of Copyrighted Work to
negotiate where computer-generated works fit within contemporary con-
ceptions of copyright. These are discussions that must be revisited by
scholars in more judicially minded fields, as well as by social scientists
 76 Publishing and Book Culture

and humanities scholars with eyes that are trained to navigate through
murky grey areas.
Finally, Section 4’s argument for the consideration of NLG systems as
social agents may be extended to systems that produce computer-generated
music, visual art, or other kinds of output typically regarded as manifesta-
tions of human communicational intention and/or creativity. These dis-
cussions have already begun, but may benefit from the incorporation of that
which has been argued here.
We have long been at a point at which systems are generating texts that
are indistinguishable from those written by humans. In recent years, how-
ever, algorithmic authorship has become ever more pervasive in the every-
day lives of ordinary readers. Computer-generated news articles and
business reports hold increasing power to inform readers’ understandings
of themselves and the world around them. Whether a personalised audit
report (with the help of Narrative Science), a bespoke book-length report
about Facial Tissue Stock Excluding Toweling, Napkin, and Toilet Paper (à la
Philip Parker), or a 50,000-word ‘novel’ produced in observation of
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

National Novel Generation Month, computer-generated texts have innu-

merable forms and uses. We are already reading computer-generated texts.
We are only just beginning to understand what that means.
 References
Abrahams, M. (2008). Automatic Writing: Further Volumes of Philip M
Parker. The Guardian. www.theguardian.com/education/2008/feb/
05/highereducation.research1 [accessed 30 April 2018].
Ackoff, R. (1989). From Data to Wisdom. Journal of Applied Systems
Analysis, 16(1), 3–9.
ALAMO (n.d.). Alamo. www.alamo.free.fr [accessed 2 October 2018].
Amazon Customer. (2004, 12 February). Review of The Official Patient’s
Sourcebook on Acne Rosacea, by James N. Parker M.D. Amazon,
www.amazon.com/product-reviews/0597832129 [accessed 27
November 2016].
Anderson, B. (2006). Imagined Communities: Reflections on the Origin and
Spread of Nationalism. London: Verso.
Anon. (1841a). From the Poughkeepsie Journal. – Hydraulic Poetry. Green
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Bay Republican, 1.
(1841b). Machine Poetry. The Cincinnati Enquirer, 2.
(1844). The New Patent Novel Writer. Punch, 6, 268.
(1845). A Latin Hexameter Machine. The Athenæum, 921, 621.
(1846). The Life of Wollaston. In The British Quarterly Review, vol. 4.
London: Jacking & Walford, 81–115.
(2010). Robot with Mechanical Brain Thinks Up Story Plots (March 1931).
Modern Mechanix. http://blog.modernmechanix.com/robot-with-
mechanical-brain-thinks-up-story-plots [accessed 21 September 2017].
(2016a). Oakville A’s 11U AAA (Mosquito) Outhits Vaughan in 7–3 Defeat.
GameChanger. https://gc.com/game-57ca09c8348c02c25400003c/recap-
story [accessed 24 March 2017].
(2016b). Recap Stories Now Available in the App! GameChanger Blog.
https://blog.gc.com/2016/07/18/recap-stories-app [accessed 26
November 2016].
 78 References

(2017). Majority of New Mothers in Wolverhampton Are Unmarried.

Express & Star. www.expressandstar.com/news/local-hubs/wolver
hampton/2017/11/29/majority-of-new-mothers-in-wolverhamp
ton-are-unmarried [accessed 31 January 2019].
(2019). Personalizing Commentary for Credit Insurance Allocation.
Yseop. www.yseop.com/case-studies/personalizing-commentary-
for-credit-insurance-allocation [accessed 12 March 2020].
Balpe, J-P. (1995). Pour une literature informatique: Un manifeste . . . . In A.
Vuillemin and M. Lenoble, eds., Littérature et informatique: La literature
générée par ordinateur. Arras: Artois Presses Université, pp. 19–32.
Barthes, R. (1977). The Death of the Author. In S. Heath, trans., Image,
Music, Text. London: Fontana Press, pp. 142–8.
Beer, G. (2014). The Reader as Author. Authorship, 3(1). http://dx.doi
.org/10.21825/aj.v3i1.1066 [accessed 3 May 2017].
Boden, M. A. (2010). Creativity and Art: Three Roads to Surprise. Oxford:
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Oxford University Press.

Bök, C. (2002). The Piecemeal Bard is Deconstructed: Notes Toward a
Potential Robopoetics. Object 10: Cyberpoetics, 10–8. www.ubu.com/
papers/object.html [accessed 28 October 2016].
Bolter, J. D. (1984). Turing’s Man: Western Culture in the Computer Age.
London: Duckworth.
(1991). Writing Space: The Computer, Hypertext, and the History of
Writing. Hillsdale, NJ: Laurence Erlbaum Associates.
Booth, W. C. (1961). The Rhetoric of Fiction. Chicago: University of
Chicago Press.
Borges, J. L. (1999). Pierre Menard, Author of the Quixote. In A. Hurley,
trans., Collected Fictions. London: Penguin, pp. 88–95.
Bosker, B. (2013). Philip Parker’s Trick For Authoring Over 1 Million
Books: Don’t Write. HuffPost. www.huffingtonpost.co.uk/entry/phi
lip-parker-books_n_2648820 [accessed 30 April 2018].
 References 79

Bowman, S. R., Vilnis, L., Vinyals, O., et al. (2016). Generating Sentences
from a Continuous Space. In Proceedings of the 20th SIGNLL
Conference on Computational Natural Language Learning. Berlin:
Association for Computational Linguistics, pp. 10–21.
Bryson, J. J. (2018). Patiency Is Not a Virtue: the Design of Intelligent Systems
and Systems of Ethics. Ethics and Information Technology, 20, 15–26.
BTN.com Staff. (2012). First Quarter Recap: UNLV 3, Minnesota 0. Big
Ten Network. http://btn.com/2012/08/30/first-quarter-track-minne
sota-at-unlv [accessed 16 April 2018].
(2013). Wisconsin Beats Michigan, 68–59. Big Ten Network. http://btn
.com/2013/03/15/track-no-4-wisconsin-vs-no-5-michigan [accessed
26 November 2016].
Casebourne, I. (1996). The Grandmother Program: A Hybrid System for
Automated Story Generation. In Creativity and Cognition 1996
Conference Proceedings. Loughborough: The Creativity and
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Cognition Studios, pp. 146–55.

Clarke, A. C. (2000). The Steam-Powered Word Processor. In The
Collected Stories of Arthur C. Clarke. New York: Tom Doherty
Associates, pp. 930–4 [first published in Analog, January 1986].
Cohen, N. (2008). He Wrote 200,000 Books (But Computers Did Some of
the Work). The New York Times. www.nytimes.com/2008/04/14/
business/media/14link.html [accessed 30 April 2018].
Cook, W. W. (1928). Plotto: A New Method of Plot Suggestion for Writers of
Creative Fiction. Battle Creek: Ellis Publishing Company.
Cowan, S. (2012). The Growth of Public Literacy in Eighteenth-Century
England. Unpublished doctoral thesis, University of London.
https://discovery.ucl.ac.uk/id/eprint/10019999/2/__d6_Shared
$_SUPP_Library_User%20Services_Circulation_Inter-Library%20Lo
ans_IOE%20ETHOS_EThOS%20-%20Redacted%20theses_COWAN,
%20S_Redacted.pdf [accessed 21 November 2020].
 80 References

Dahl, R. (1954). The Great Automatic Grammatisator. In Someone Like

You. New York: Alfred A. Knopf, pp. 250–76.
Darnton, R. (1982). What Is the History of Books? Daedalus, 111(3), 65–83.
Das, R., and Pavlíčková, T. (2013). Is There an Author Behind This Text?
A Literary Aesthetic Driven Approach to Interactive Media. New
Media & Society, 16(3), 381–97.
Deazley, R. (2008). Commentary on Millar v. Taylor (1769). Primary
Sources on Copyright (1450–1900), ed. by Lionel Bently and Martin
Kretschmer. www.copyrighthistory.org/cam/tools/request/
showRecord?id=commentary_uk_1769 [accessed 20 June 2019].
Dehn, N. J. (1989). Computer Story-Writing: The Role of Reconstructive and
Dynamic Memory. New Haven: Yale University Department of
Computer Science.
Dennett, D. (1971). Intentional Systems. The Journal of Philosophy, 68(4),
87–106.
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

(1987). The Intentional Stance. Cambridge, MA: The MIT Press.

Dewdney, A. K. (1985). Artificial Insanity: When a Schizophrenic
Program Meets a Computerized Analyst. Scientific American,
252(1), 10–3.
Downey, N. (2015, 8 March). Review of The Official Patient’s Sourcebook on
Acne Rosacea, by James N. Parker M.D. Amazon. www.amazon.com/
product-reviews/0597832129 [accessed 27 November 2016].
Eisenstein, E. L. (1979). The Printing Press as an Agent of Change.
Cambridge: Cambridge University Press.
Eureka AHRC Project. (n.d.) Alfred Gillett Trust. https://alfredgilletttrust
.org/collections/latin-verse-machine [accessed 7 May 2019].
Eve, M. P. (2017). The Great Automatic Grammatizator: Writing, Labour,
Computers. Critical Quarterly, 59(3), 39–54.
Febvre, L., and Martin, H-J. (1997). The Coming of the Book: The Impact of
Printing, 1450–1800, trans. by David Gerard. London: Verso.
 References 81

Fischer, K., and Bateman, J. A. (2006). Keeping the Initiative: An

Empirically-Motivated Approach to Predicting User-Initiated
Dialogue Contribution in HCI. In Proceedings of the 11th
Conference of the European Chapter of the Association for
Computational Linguistics. Trento: Association for Computational
Linguistics, pp. 185–92.
Fish, S. (1980). Is There a Text in This Class? The Authority of Interpretive
Communities. Cambridge, MA: Harvard University Press.
Foucault, M. (1998). What Is an Author? In J. D. Faubion, ed. and R.
Hurley et al., trans., Aesthetics, Method, and Epistemology. New York:
The New Press, pp. 205–22.
Gauchet, M. (2000). A New Age of Personality: An Essay on the
Psychology of our Times. Thesis Eleven, 60, 23–41.
Goldschmidt, E. P. (1943). Medieval Texts and Their First Appearance in
Print. London: The Bibliographical Society.
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Goody, J., and Watt, I. (1963). The Consequences of Literacy. Comparative

Studies in Society and History, 5(3), 304–45.
Gregory, J., and Miller, S. (1998). Science in Public: Communication, Culture,
and Credibility. New York: Basic Books.
Hall, J. D. (2007). Popular Prosody: Spectacle and the Politics of
Victorian Versification. Nineteenth-Century Literature, 62(2),
222–49.
(2017). Nineteenth-Century Verse and Technology: Machines of Meter.
Cham: Palgrave Macmillan.
Hartman, C. O. (1996). Virtual Muse: Experiments in Computer Poetry.
Hanover, NH: Wesleyan University Press.
Havelock, E. A. (1980). The Coming of Literate Communication to
Western Culture. Literacy and the Future of Print, 30(1), 90–8.
Hayles, N. K. (2005). My Mother Was a Computer: Digital Subjects and
Literary Texts. Chicago: University of Chicago Press.
 82 References

Henrickson, L. (2018a). Computer-Generated Fiction in a Literary

Lineage: Breaking the Hermeneutic Contract. Logos, 29(2–3),
54–63.
(2019a). Natural Language Generation: Negotiating Text Production in
Our Digital Humanity. Proceedings of the Digital Humanities Congress
2018. Sheffield: University of Sheffield. www.dhi.ac.uk/openbook/
chapter/dhc2018-henrickson [accessed 28 February 2020].
(2018b). Tool vs. Agent: Attributing Agency to Natural Language
Generation Systems. Digital Creativity, 29(2–3), 182–90.
(2019b). ‘Towards a New Sociology of the Text: The Hermeneutics of
Algorithmic Authorship’ Empirical Studies. Loughborough University
Research Repository. https://doi.org/10.17028/rd.lboro.c.4663709
[accessed 28 February 2020].
(2019c). ‘Towards a New Sociology of the Text: The Concept of the
Author in the Digital Age’ Focus Group Transcriptions.
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Loughborough University Research Repository. https://doi.org/

10.17028/rd.lboro.7923944 [accessed 4 February 2020].
(2020). Authorship in Computer-Generated Texts. Oxford Research
Encyclopedia of Literature. https://doi.org/10.1093/acrefore/
9780190201098.013.1226 [accessed 18 August 2020].
Herman, D. (2008). Narrative Theory and the Intentional Stance. Partial
Answers: Journal of Literature and the History of Ideas, 6(2), 233–60.
Hill, W. A. (1919). Ten Million Photoplay Plots. Los Angeles: Feature
Photodrama.
Hofstadter, D. R. (1980). Gödel, Escher, Bach: An Eternal Golden Braid.
London: Penguin.
Houston, R. A. (1993). Literacy, Education and the Culture of Print in
Enlightenment Edinburgh. History, 78(254), 373–92.
hugovk. (2014). 50,000 Meows. GitHub. https://github.com/dariusk/
NaNoGenMo-2014/issues/50 [accessed 29 August 2018].
 References 83

ICON Group International. (n.d.) www.icongrouponline.com [accessed 25

November 2016].
Immersive Automation. (n.d.) http://immersiveautomation.com [accessed 1
October 2017].
Johnson, D. G. (2006). Computer Systems: Moral Entities But Not Moral
Agents. Ethics and Information Technology, 8, 195–204.
Karsdorp, F., Manjavacas, E., Burtenshaw, B., Kestemont, M., and
Stokhuyzen, B. (2017). AsiBot. https://asibot.nl [accessed 17 December
2018].
Kazemi, D. (n.d.) NaNoGenMo. GitHub. https://nanogenmo.github.io
[accessed 29 August 2018].
Kirschenbaum, M. (2015). What Is an @uthor? Los Angeles Review of Books.
http://lareviewofbooks.org/essay/uthor [accessed 7 February 2015].
Klein, S., Aeschlimann, J. F., Balsiger, D. F., et al. (1973). Automatic Novel
Writing: A Status Report (Technical Report 186). Madison: The
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

University of Wisconsin.
Koolhof, K. (2017). Ronald Giphart experimenteert met literaire robot. AD.
nl. www.ad.nl/wetenschap/ronald-giphart-experimenteert-met-litera
ire-robot~a5a3cb9f [accessed 1 October 2017].
Krittman, D., Matthews, P., and Glascott, M. G. (2015). Innovation Ushers in
the Modern Era of Compliance. Deloitte. www2.deloitte.com/content/
dam/Deloitte/us/Documents/finance/us-fas-how-natural-language-
is-changing-the-game-deloitte-only.pdf [accessed 15 March 2018].
Lake, R. W. (2017). Big Data, Urban Governance, and the Ontological
Politics of Hyperindividualism. Big Data & Society. https://doi.org/
10.1177/2053951716682537 [accessed 1 July 2019].
Latour, B. (2005). Reassembling the Social: An Introduction to Actor-Network-
Theory. Oxford: Oxford University Press.
Lebowitz, M. (1985). Story-Telling as Planning and Learning. Poetics, 14,
483–502.
 84 References

Lee, P. (2016). Learning From Tay’s Introduction. Official Microsoft Blog.

http://blogs.microsoft.com/blog/2016/03/25/learning-tays-intro
duction [accessed 8 November 2016].
Leppänen, L., Munezero, M., Sirén-Heikel, S., Granroth-Wilding, M., &
Toivonen, H. (2017). Finding and Expressing News From Structured
Data. In Proceedings of the 21st International Academic Mindtrek
Conference. Tampere: Association for Computing Machinery, pp.
174–83.
Lescure, J. (1973). La méthode S + 7 (cas particulier de la méthode M ± n).
In La littérature potentielle (Créations, Re-créations, Récréations. Paris:
Éditions Gallimard, pp. 143–8.
Malm, P. (2020). The Language Effect: Why AI-Powered Copywriting Is a
Marketer’s (New) Best Friend. London: Known Publishing.
Manjavacas, E., Karsdorp, F., Burtenshaw, B., and Kestemont, M. (2017).
Synthetic Literature: Writing Science Fiction in a Co-Creative
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Process. In Proceedings of the INLG 2017 Workshop on Computational

Creativity and Natural Language Generation. Santiago de Compostela:
Association for Computational Linguistics, pp. 29–37.
McDonald, D. D. (1986). Natural Language Generation: Complexities and
Techniques [internal memo]. University of Massachusetts Department
of Computer and Information Science. Amherst: University of
Massachusetts.
McGuinness, R. (2014). Meet the Robots Writing Your News Articles: The
Rise of Automated Journalism. Metro. http://metro.co.uk/2014/07/
10/meet-the-robots-writing-your-news-articles-the-rise-of-auto
mated-journalism-4792284 [accessed 30 April 2018].
McKenzie, D. F. (1999). Bibliography and the Sociology of Texts.
Cambridge: Cambridge University Press.
McLuhan, M. (1994). Understanding Media: The Extensions of Man.
Cambridge, MA: Massachusetts Institute of Technology.
 References 85

McLuhan, M., Fiore, Q., and Agel, J. (1967). The Medium is the Massage:
An Inventory of Effects. New York: Bantam Books.
Meehan, J. R. (1976). The Metanovel: Writing Stories by Computer.
Unpublished Ph.D thesis, Yale University. www.semanticscholar
.org/paper/The-Metanovel%3A-Writing-Stories-by-Computer-
Meehan/35f03721ecef2a7315a8d85d02bacaf00660a3fb [accessed 21
November 2020].
Menabrea, L. F. (1843). Sketch of the Analytical Engine Invented by
Charles Babbage, Esq. In A. A. Lovelace, trans., R. Taylor, ed.,
Scientific Memoirs, Selected from the Transactions of Foreign
Academies of Science and Learned Societies, and from Foreign Journals,
Vol. III. London: Richard and John E. Taylor, pp. 666–731.
Method and Apparatus for Automated Authoring and Marketing (US
7266767 B2). (2006). Google Patents. https://patents.google.com/
patent/US7266767 [accessed 9 October 2017].
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Michael, M. (2014). How to Understand Mundane Technology: New Ways

of Thinking about Human-Technology Relations. In J. R. Dakers,
ed., Defining Technological Literacy: Towards an Epistemological
Framework, 2nd ed. New York: Palgrave Macmillan, pp. 41–56.
Mittelstadt, B. D., Allo, P., Taddeo, M., Wachter, S., and Floridi, L. (2016).
The Ethics of Algorithms: Mapping the Debate. Big Data & Society,
3(2), 1–21.
Montal, T., and Reich, Z. (2017). I, Robot. You, Journalist. Who is the
Author? Digital Journalism, 5(7), 829–49.
Montfort, N. (2005). Twisty Little Passages: An Approach to Interactive
Fiction. Cambridge, MA: The MIT Press.
Montfort, N., and Fedorova, N. (2012). Small-Scale Systems and
Computational Creativity. In Proceedings of the Third International
Conference on Computational Creativity. Dublin: Association for
Computational Creativity, pp. 82–6.
 86 References

Motta, M. (2018). The Dynamics and Political Implications of

Anti-Intellectualism in the United States. American Politics Research,
43(3), 465–98.
Murray, S. (2018a). The Digital Literary Sphere: Reading, Writing, and Selling
Books in the Internet Era. Baltimore: Johns Hopkins University Press.
(2018b). Reading Online: Updating the State of the Discipline. Book
History, 21, 370–96.
Nass, C., and Moon, Y. (2000). Machines and Mindlessness: Social
Responses to Computers. Journal of Social Issues, 56(1), 81–103.
Nass, C., Steuer, J., & Tauber, E. R. (1994). Computers are Social Actors. In
CHI ‘94 Proceedings of the SIGCHI Conference on Human Factors in
Computing Systems. Boston: Association for Computing Machinery,
pp. 72–8.
Natale, S. (in press). Deceitful Media. Oxford: Oxford University Press.
National Commission on New Technological Uses of Copyrighted Works.
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

(1979). Final Report of the National Commission of New Technological

Uses of Copyright Works. Washington: Library of Congress.
Nehamas, A. (1986). What an Author Is. The Journal of Philosophy, 83(11),
685–91.
Nuttall, P. A. (1845). The Eureka. In J. Holmes Agnew (ed.), The Eclectic
Magazine of Foreign Literature, Science, and Art. New York: Leavitt,
Trow, & Co., pp. 140–1.
Parker, P. M. (2006a). The 2007 Report on Facial Tissue Stock Excluding
Toweling, Napkin, and Toilet Paper: World Market Segmentation by
City. Las Vegas: Icon Group International.
(2006b). The 2007–2012 Outlook for Instant Chocolate Milk, Weight Control
Products, Whole Milk Powder, Malted Milk Powder, and Other Dry Milk
Products Shipped in Consumer Packages Weighing 3 Pounds or Less
Excluding Nonfat Dry Milk and Infants’ Formula in Japan. Las Vegas:
Icon Group International.
 References 87

(2007). Webster’s Swedish to English Crossword Puzzles: Level 1. Las

Vegas: Icon Group International.
Parrish, A. (2015). Our Arrival: A Novel. http://s3.amazonaws.com/apar
rish/our-arrival.pdf [accessed 29 August 2018].
Pavlíčková, T. (2013). Bringing the Author Back Into the Audience Research:
A Hermeneutical Perspective on the Audience’s Understanding of the
Author. The Communication Review, 16(1–2), 31–9.
Peter, J. (1677). Artificial Versifying | A New Way to Make Latin Verſes.
Whereby Any one of Ordinary Capacity, that only knows the A.B.C. And
can count 9 (though he understands not One Word of Latin, or what a
Verſe means) may be plainly taught (and in as little time, as this is
reading over) how to make Thouſands of Hexameter and Pentameter
Verſes which ſhall be True Latine & True Verſe, and Good Senſe., 2nd
ed. London: Kings-Head at Sweetings-Alley end, next House to the
Royal Exchange in Corn-hill.
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Pinch, T. J., and Bijker, W. E. (1984). The Social Construction of Facts and
Artefacts: Or How the Sociology of Science and the Sociology of
Technology Might Benefit Each Other. Social Studies of Science, 14,
399–441.
Podolny, S. (2015). If an Algorithm Wrote This, How Would You Even
Know? The New York Times. www.nytimes.com/2015/03/08/opi
nion/sunday/if-an-algorithm-wrote-this-how-would-you-even-
know.html [accessed 30 April 2018].
Polti, G. (1895). Les trente-six situations dramatiques. Paris: Mercure de
France.
Portela, M. (2018). Writing under Constraint of the Regime of Computation.
In J. Tabbi, ed., The Bloomsbury Handbook of Electronic Literature.
London: Bloomsbury, pp. 181–200.
Powers, T. M. (2013). On the Moral Agency of Computers. Topoi, 32(2),
227–36.
 88 References

Prescott, T. J. (2017). Robots Are Not Just Tools. Connection Science, 29(2),
142–9.
Racter [Chamberlain, W., and Etter, T. ]. (1984). The Policeman’s Beard Is
Half Constructed. New York: Warner Software/Warner Books.
Radford, A., Wu, J., Amodei, D., et al. (2019). Better Language Models and
Their Implications. OpenAI. https://openai.com/blog/better-lan
guage-models [accessed 20 August 2020].
Ray Murray, P., and Squires, C. (2013). The Digital Publishing
Communications Circuit. Book 2.0, 3(1), 3–23.
Reeves, B., and Nass, C. (1996). The Media Equation: How People Treat
Computers, Television, and New Media Like Real People and Places.
Cambridge: Cambridge University Press.
Regan, D. (2015). The Cover of The Sun Also Rises. GitHub. http://
alsorises.org [accessed 29 August 2018].
Roberts, S. (2017). Christopher Strachey’s Nineteen-Fifties Love Machine. The
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

New Yorker. www.newyorker.com/tech/annals-of-technology/christo

pher-stracheys-nineteen-fifties-love-machine [accessed 11 March 2019].
Ryan, J. (2017). Grimes’ Fairy Tales: A 1960s Story Generator. In
Interactive Storytelling: 10th International Conference on Interactive
Digital Storytelling. Cham: Springer, pp. 89–103.
Saenger, P. (1999). Reading in the Later Middle Ages. In G. Cavallo and R.
Chartier, eds., L. G. Cochrane, trans., A History of Reading in the
West. Cambridge: Polity Press, pp. 120–48.
Schoff Curtin, R. (2017). The Transactional Origins of Authors’
Copyright. The Columbia Journal of Law and the Arts, 40(2), 175–235.
Searle, J. R. (1980). Minds, Brains, and Programs. The Behavioral and Brain
Sciences, 3(3) 417–57.
Shillingsburg, P. L. (2006). From Gutenberg to Google: Electronic
Representations of Literary Texts. Cambridge: Cambridge University
Press.
 References 89

Skains, R. L. (2019). Digital Authorship. Cambridge: Cambridge University

Press.
Smiggy. (2002, 19 December). Follow Your Nose, Couldn’t Put It Down!
[Customer Review]. Review of The 2007 Report on Facial Tissue Stock
Excluding Toweling, Napkin, and Toilet Paper: World Market
Segmentation by City, by Philip M. Parker. Amazon. www.amazon.co
.uk/gp/customer-reviews/RCNPTIBJJNAQV [accessed 30 April
2018].
Spence, P. R., Edwards, A., Edwards, C., and Jin, X. (2019). ‘The Bot Predicted
Rain, Grab an Umbrella’: Few Perceived Differences in Communication
Quality of a Weather Twitterbot Versus Professional and Amateur
Meteorologists. Behaviour & Information Technology, 38(1), 101–9.
Stahl, B. C. (2004). Information, Ethics, and Computers: The Problem of
Autonomous Moral Agents. Minds and Machines, 14(1), 67–83.
Tang, J., Yang, Y., Carton, S., Zhang, M., and Mei, Q. (2016). Context-
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Aware Natural Language Generation with Recurrent Neural

Networks. Cornell University Library (arXiv) https://arxiv.org/abs/
1611.09900 [accessed 25 June 2017].
Teigen, P. M. (1987). A Prolegomenon to the Interpretation of The Printing
Press as an Agent of Change. In P. F. McNally, ed., The Advent of Printing:
Historians of Science Respond to Elizabeth Eisenstein’s The Printing Press as
an Agent of Change. Montreal: McGill University, pp. 8–14.
Tenen, D. (2017). Plain Text: The Poetics of Computation. Stanford:
Stanford University Press.
Tomasula, S. (2018). Our Tools Make Us (and Our Literature) Post. In J.
Tabbi, ed., The Bloomsbury Handbook of Electronic Literature. London:
Bloomsbury, pp. 39–58.
Turner, S. R. (1993). MINSTREL: A Computer Model of Creativity and
Storytelling. Unpublished doctoral thesis, University of California Los
Angeles. www.semanticscholar.org/paper/Minstrel%3A-a-compu
 90 References

ter-model-of-creativity-and-Turner/745b24c90f089339f7e9e2209d4a
cebfb2f1ec82 [accessed 21 November 2020].
Urbs Media. (n.d.). www.urbsmedia.com [accessed 25 February 2018].
Valtteri (n.d.) [archived by Internet Archive]. https://web.archive.org/web/
20190120010217/https://www.vaalibotti.fi [accessed 18 August 2020].
van der Weel, A. (2001). The Communications Circuit Revisited. Jaarboek
voor Nederlandse boekgeschiedenis, 8, 13–25.
(2011a). Changing Our Textual Minds: Towards a Digital Order of
Knowledge. Manchester: Manchester University Press.
(2011b). Our Textual Future. Logos, 22(3), 44–52.
(2014). From an Ownership to an Access Economy of Publishing. Logos,
25(2), 39–46.
(2015). Appropriation: Towards a Sociotechnical History of Authorship.
Authorship, 4(2). https://doi.org/10.21825/aj.v4i2.1438 [accessed 1 July
2019].
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

Vuillemin, A., and Lenoble M., eds. (1995). Littérature et informatique: La

literature générée par ordinateur. Arras: Artois Presses Université.
Weizenbaum, J. (1966). ELIZA – A Computer Program for the Study of
Natural Language Communication Between Man and Machine.
Communications of the ACM, 9(1), 36–45.
Whalen, Z. (2018a). A Python Script that Writes 800-page Children’s
Books. Zach Whalen. www.zachwhalen.net/posts/a-python-script-
that-writes-800-page-childrens-books [accessed 30 August 2018].
(2018b). The Several Houses of Brian, Spencer, Liam, Victoria, Brayden,
Vincent, and Alex. GitHub. https://github.com/zachwhalen/nngm17
[accessed 28 August 2018].
(2019). The Many Authors of The Several Houses of Brian, Spencer,
Liam, Victoria, Brayden, Vincent, and Alex: Authorship, Agency, and
Appropriation. Journal of Creative Writing Studies, 4(1) https://scho
larworks.rit.edu/jcws/vol4/iss1/6. [accessed 20 July 2020].
 References 91

Wiener, N. (1950). The Human Use of Human Beings: Cybernetics and

Society. London: Eyre and Spottiswoode.
Wilson, A. (2004). Foucault on the ‘Question of the Author’: A Critical
Exegesis. The Modern Language Review, 99(2), 339–63.
Winograd, T. (1971). Procedures as a Representation for Data in a Computer
Program for Understanding Natural Language. Unpublished doctoral
thesis, Massachusetts Institute of Technology. https://dspace.mit
.edu/handle/1721.1/7095 [accessed 21 November 2020].
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press
 Acknowledgements
This Element would hardly have been possible without the guidance and
insight of, among many others: Drs Wim Van Mierlo, Clare Hutton,
Simone Natale, and Laura Dietz, as well as Professors Adriaan van der
Weel and Jenny Fry. Thanks are also due to Loughborough University’s
Digital Humanities Research Group and the Martin Hall crew. For the less
sexy but more comprehensive version of this book (i.e. my doctoral thesis),
see https://doi.org/10.26174/thesis.lboro.10596404.
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press
 Publishing and Book Culture
SERIES EDITOR
Samantha Rayner
University College London
Samantha Rayner is a Reader in UCL’s Department of
Information Studies. She is also Director of UCL’s Centre for
Publishing, co-Director of the Bloomsbury CHAPTER
(Communication History, Authorship, Publishing, Textual
Editing and Reading) and co-editor of the Academic Book of
the Future BOOC (Book as Open Online Content) with UCL
Press.
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

ASSOCIATE EDITOR
Leah Tether
University of Bristol
Leah Tether is Professor of Medieval Literature and Publishing
at the University of Bristol. With an academic background in
medieval French and English literature and a professional
background in trade publishing, Leah has combined her
expertise and developed an international research profile in
book and publishing history from manuscript to digital.
 ADVISORY BOARD

Simone Murray, Monash Alan Staton, Booksellers

University Association
Claire Squires, University of Angus Phillips, Oxford
Stirling International Centre for
Andrew Nash, University of Publishing
London Richard Fisher, Yale University
Leslie Howsam, Ryerson Press
University John Maxwell, Simon Fraser
David Finkelstein, University of University
Edinburgh Shafquat Towheed, The Open
Alexis Weedon, University of University
Bedfordshire Jen McCall, Emerald Publishing
https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

ABOUT THE SERIES

This series aims to fill the demand for easily accessible, quality texts
available for teaching and research in the diverse and dynamic fields
of Publishing and Book Culture. Rigorously researched and
peer-reviewed Elements will be published under themes, or
‘Gatherings’. These Elements should be the first check point for
researchers or students working on that area of publishing and book
trade history and practice: we hope that, situated so logically at
Cambridge University Press, where academic publishing in the UK
began, it will develop to create an unrivalled space where these
histories and practices can be investigated and preserved.
 Publishing and Book Culture
Digital Literary Culture
Gathering Editor: Laura Dietz
Laura Dietz is a Senior Lecturer in Writing and Publishing in
the Cambridge School of Creative Industries at Anglia Ruskin
University. She writes novels and studies novels, publishing
fiction alongside research on topics such as e-novel readership,
the digital short story, online literary magazines, and the
changing definition of authorship in the digital era.

ELEMENTS IN THE GATHERING

https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press

The Network Turn: Changing Perspectives in the Humanities

Ruth Ahnert, Sebastian E. Ahnert, Catherine Nicole Coleman and Scott
B. Weingart

Reading Computer-Generated Texts

Leah Henrickson

A full series listing is available at: www.cambridge.org/EPBC

https://doi.org/10.1017/9781108906463 Published online by Cambridge University Press