1

This is an unpublished working paper that has not undergone peer review.

08/21/2023

Misinformation inoculations must be boosted by

accuracy prompts to improve judgments of truth
Gordon Pennycook1,2*, Adam J. Berinsky3, Puneet Bhargava4,5, Rocky Cole6, Beth Goldberg6,

Stephan Lewandowsky7,8,9, & David G. Rand10,11,12

Department of Psychology, Cornell University, 2Hill/Levene Schools of Business, University of Regina,

1

Department of Political Science, Massachusetts Institute of Technology, 4Department of Psychology, University of

3

Pennsylvania, 5Department of Marketing, The Wharton School, University of Pennsylvania, 6Google, 7School of
Psychological Science, University of Bristol, 8Department of Psychology, University of Potsdam, 9School of
Psychological Science, University of Western Australia, 10Sloan School of Management, Massachusetts Institute of
Technology 11Institute for Data, Systems, and Society, Massachusetts Institute of Technology, 12Department of
Brain and Cognitive Sciences, Massachusetts Institute of Technology

*Corresponding author: gordon.pennycook@cornell.edu, Cornell University, 211 Tower Rd, Ithaca, NY, United
States, 14850

Misinformation remains a serious problem and continues to be a major focus of intervention

efforts. Psychological inoculation - a popular intervention approach wherein people are
taught to identify manipulation techniques - is being adopted at scale around the globe by
technology companies in an effort to combat misinformation. Yet the efficacy of this
approach for increasing belief accuracy remains unclear, as prior work has largely focused
on technique identification - rather than accuracy judgments - using synthetic materials that
do not contain claims of truth or falsity. To address this issue, we conducted 5 studies with
7,286 online participants using a set of news headlines based on real-world false and true
content in which we systematically varied the presence or absence of emotional
manipulation. Although an emotional manipulation inoculation video did help participants
identify emotional manipulation (replicating past work), there was no carry-over effect to
improving participants’ ability to tell truth from falsehood (i.e. no effect on truth
discernment). Encouragingly, however, when the inoculation was paired with an accuracy
prompt - i.e., an intervention intended to draw people’s attention to the concept of accuracy
- the combined intervention did successfully improve truth discernment. These results
generate new insights regarding inoculation, and provide evidence for a key synergy between
two popular psychological interventions against misinformation.
 2

Global concerns about the spread of misinformation and its toxic effects on democracy have led
to a surge of interest in potential mitigation measures (Athey et al., 2023; Bak-Coleman et al.,
2022; Ecker et al., 2022; Kozyreva et al., 2020, 2022; Pennycook & Rand, 2021). One of the
most common approaches is to undermine the influence of falsehoods by providing
countervailing information. For example, to ensure that people do not believe misinformation
(defined here as information that is false or misleading, whatever the intent), research has
focused on assessing the efficacy of fact-checks, corrections, and “debunks” (Chan et al., 2017;
Lewandowsky et al., 2012; Nieminen & Rapeli, 2019; Porter & Wood, 2021). Although these
approaches show promise – corrections do tend to undermine the effect of misinformation and
rarely backfire (Wood & Porter, 2019) – such an approach is a loss-mitigation strategy at best.
Corrections necessarily occur after misinformation has already spread online, and thus should
only be one element of a broader strategy.

Psychological inoculation

Researchers and practitioners have therefore also focused on strategies that attempt to get ahead
of misinformation and limit its potential influence before people are exposed. Recent research
demonstrates that “prebunking” misinformation by providing corrective information in
anticipation of falsehoods can protect people against being misled (Cook et al., 2017;
Lewandowsky & van der Linden, 2021; Traberg et al., 2023). One promising prebunking
strategy that has been gaining popularity is attitudinal or psychological “inoculation”. The goal
of inoculation is to provide individuals with knowledge or abilities that help them spot
misinformation by exposing them to a weakened form of a misleading technique (or set of
techniques) (Roozenbeek et al., 2022; Roozenbeek & van der Linden, 2022; Traberg et al.,
2022). Critically, unlike other prebunking techniques, inoculation techniques are not intended to
be specific to unique falsehoods, but rather are intended to boost people's general ability to
identify misleading techniques that are thought to be common to misinformation (Compton,
2013; Compton et al., 2021).

To illustrate, a recent set of experiments have tested a scalable version of psychological

inoculation. Roozenbeek et al. (2022) developed a set of brief videos (less than two minutes
each) that teach users about manipulation techniques that are thought to be statistical markers of
misinformation (namely emotional language, ad hominem attacks, false dichotomies,
incoherence, and scapegoating). To offer an example, the emotional manipulation inoculation
video informs users about how emotional language (e.g., words like “disgusting” and
“horrifying”) can be used to manipulate users into paying attention to content online (all videos
are accessible here, https://inoculation.science/). These videos, generally speaking, improved
participants’ ability to discern whether there was manipulation in the synthetic tweets that the
researchers created to either contain the technique or not – with the emotional language
inoculation having one of the clearest effects on the ability to discern between manipulative and
neutral content.
 3

Two of the inoculation videos were also tested in a quasi-field experiment on YouTube. For this,
users were presented an inoculation video in the standard course of their YouTube usage as an ad
and then were subsequently asked to correctly identify the manipulation technique in a multiple
choice question. The researchers found those who viewed the inoculative video were 5-6% more
likely to correctly identify a manipulation technique relative to a control group, suggesting that a
short video on social media may be sufficient to teach people how to identify misinformation
techniques (Roozenbeek et al., 2022). Efforts by technology companies to scale this approach are
already underway (Jigsaw, 2023; Klepper, 2023; Mukherjee et al., 2022).

Although inoculation thus shows considerable promise as a scalable approach, research on

inoculation has focused on demonstrating people’s ability to detect when they might be misled
by boosting their ability to recognize misleading or deceptive argumentation (Roozenbeek et al.,
2022; Traberg et al., 2023). This is a valuable skill, in particular because it provides protection
against poor argumentation irrespective of the specific content. For example, the use of highly
emotive or inflammatory language may signal manipulative intent irrespective of whether the
message is entirely false, contains seeds of truth, or is accurate.

However, it remains unclear whether inoculation also makes people better at discerning the
accuracy of claims – that is, whether inoculation increases the ability to differentiate true from
false information. Past work has primarily tested the efficacy of inoculation by using content that
contains the targeted technique (such as emotional manipulation), but that does not necessarily
contain specific truth claims (or claims that are evaluable as true or false). For example,
Roozenbeek et al. (2022) used the following as an example of emotional language manipulation:
“What this airline did for its passengers will make you tear up - So heartwarming” (with no
additional context or information about the airline or what they actually did). Items such as this
are intended to be “pure” examples of the technique; naturally, however, the missing context
would be present when such interventions are used in the real world and it is important to
ascertain whether technique identification skills extend to recognizing the truth or falsehood of
statements.

Typically, participants are also explicitly asked about the presence/absence of the technique
instead of the accuracy of the statements 1. Thus, it is unclear if boosting technique recognition
carries over to affect judgments about the truth or falsity of true and false content. Does teaching
about misleading techniques (on its own) help people detect misinformation? And if not, what
can be done to address this issue?

Inattention to accuracy

1 One exception to this is Lewandowsky and Yesilada (2021) who measured people’s ability to judge the likely
accuracy of problematic content, in addition to examining their detection ability. In this case, inoculation has also
shown promise by reducing people’s belief in the reliability of dubious information, although (as in other cases), the
test material did not have explicit truth value.
 4

Notwithstanding the promise of inoculation, there are also reasons to believe that inoculations
with a specific focus on misleading techniques (such as using emotional language) may not
necessarily carry-over to improve judgments about the accuracy of true or false content. In
particular, research has shown that people are often inattentive to accuracy on social media
(Arechar et al., 2023; Pennycook et al., 2020, 2021). For example, in many experiments when
participants are asked directly to assess whether they believe a set of true and false news
headlines, most people are actually quite good at distinguishing between them (i.e., they believe
the true news much more than the false news); however, if participants are instead asked to
indicate if they would share the news on social media, they are largely insensitive to the veracity
of the headlines (Arechar et al., 2023; Pennycook et al., 2020, 2021). In fact, the social media
context may itself be responsible for distracting people from thinking about accuracy, as simply
having participants make sharing judgments reduces their ability to tell true from false when
making accuracy judgments (Epstein et al., 2023).

One approach to overcoming this inattentiveness to accuracy is through very brief interventions
that prompt the reader to consider accuracy, known as “accuracy prompts” (or “accuracy primes”
or “accuracy nudges”). These prompts have demonstrated promise in improving people’s
attentiveness to accuracy when making judgments about what to share online (i.e., by increasing
the quality of content that they share) (Epstein et al., 2021; Pennycook et al., 2020, 2021;
Roozenbeek et al., 2021; for a meta-analysis see Pennycook & Rand, 2022). This is thought to
occur because the prompts redirect participants’ attention to the concept of accuracy, and most
participants do not wish to share content that they realize is inaccurate (Arechar et al., 2023;
Pennycook et al., 2021). Thus, simply considering accuracy reduces the (undesired) sharing of
content which participants would have been able to identify as inaccurate had they thought about
it (Lin et al., 2023)

It is therefore possible that the effects of inoculations that focus only on identifying manipulation
techniques - but do not address the problem of trying to specifically spot false claims - may not
transfer to improved misinformation detection. However, if before delivering the inoculation, the
intervention explicitly draws people’s attention to accuracy and the need to differentiate truth
from falsehood, this may induce transfer and successfully lead to improved truth discernment.
Put differently, there may be a synergy between inoculation approaches (which teach people
specific information about manipulation techniques but may not focus their attention on
detecting falsehoods) and accuracy prompt approaches (which direct attention to truth and falsity
but do not actually teach individuals anything about how to detect falsehoods).

Present work

Here we investigate whether psychological inoculation effects carry over to improved truth
discernment, as well as the potential synergy between inoculation and accuracy prompting. We
show in 3 experiments that, while inoculation reliably boosts people’s ability to detect
 5

problematic content, it does not by itself increase truth discernment. However, we show in two
further experiments that when inoculation is combined with a prompt that reminds people of the
problem of misinformation and the importance of accuracy, the combined intervention (but
neither prompt nor inoculation in isolation) boosts participants’ accuracy discernment.

To assess participants’ ability to identify inaccurate claims that systematically vary in the
presence of a manipulation technique, we developed a novel set of stimuli that were sourced
from real-world true and false news headlines but that were modified to either contain the
manipulation technique of emotional language or no such technique (see Figure 1). For this, we
took a larger corpus of 83 political headlines (43 false, 40 true) and created two versions for each
claim: One intended to elicit strong emotions from the reader by including evocative language
and one that did not include an emotional manipulation (and that was stripped of evocative
language). We then took this corpus and completed a pretest where participants (N = 982) rated
the extent to which a random subset of 15 headlines made them feel emotions (e.g., angry, sad,
worried, happy, excited, etc.) using a 6-point scale from ‘not at all’ to ‘extremely’. Participants
also rated the headlines in terms of their perceived likelihood and (Democratic v. Republican)
partisanship, and indicated whether they would share the headline on social media. From this, we
selected 32 headlines where the high emotion version elicited particularly strong emotional
reactions relative to the low emotion version (see Table 1). This procedure gave us a set of
headlines that not only differed in terms of whether they contained the emotional manipulation
technique (as in past work) but that also were pretested to be significantly different in terms of
the actual emotions that were elicited by the headlines.

This pretested set of news headlines allows us to test whether the emotional manipulation
inoculation is successful in helping people identify misinformation using real-world stimuli that
contain true or false claims. In addition, in the studies below participants were asked either about
manipulativeness or belief (accuracy) separately so that we could investigate if participants
spontaneously identify misinformation to a greater extent following the intervention.
 6

Figure 1. Example false (top) and true (bottom) headlines that were manipulated to contain an
emotional language technique (left) or no such technique (right).

Table 1. Average emotionality rating in the pretest for matched high versus low emotionality
headlines. Displayed are in-party ratings. I.e., Emotionality ratings among Democrats for
Democratic-targeted headlines and among Republicans for Republican-targeted headlines.

Target Low High Diff t-test Cohen’s

Emotionality Emotionality d

False Democratic 3.0 3.7 0.7 t(7) = 19.9, p < .001 7.02
Republican 2.8 3.3 0.5 t(7) = 10.4, p < .001 3.68
True Democratic 3.0 3.6 0.6 t(7) = 6.7, p < .001 2.37
Republican 3.1 3.6 0.5 t(7) = 4.4, p = .003 1.54
 7

Experiment 1

In testing our hypotheses, we focus on whether the inoculation treatment increases discernment
in two forms. Since we orthogonally manipulate both stimuli facticity and emotionality, we are
able to assess two versions of discernment: 1) Do users distinguish between true and false
content; i.e., are ratings of belief (manipulativeness) lower (higher) for false relative to true news
headlines?, and 2) Do users distinguish between emotionally manipulative (high emotion) and
not-emotionally manipulative (low emotion) content; i.e., are ratings of belief (manipulativeness)
lower (higher) for manipulative relative to not-manipulative news headlines?

Method

Participants

Participants were recruited from Amazon’s Mechanical Turk via Cloudresearch, which has been
shown to provide high quality participants (Douglas et al., 2023). Participants were randomized
into an emotional inoculation treatment condition or a control condition, and participants in the
control condition were further randomized to one of three control videos. In total, 1,268
participants entered the experiment. However, 33 participants indicated not having social media
and did not therefore qualify for the study and 6 participants failed an attention check at the
beginning of the study and did not continue. Finally, 3 participants were removed from the
survey because they indicated not being able to turn on their volume (which was critical for
watching the videos) and 18 participants quit the survey prior to the intervention. A further 178
participants dropped out and did not contribute data to the primary task. The fraction of
participants dropping out did not vary significantly across conditions (15. 2% in control, 14.2%
in treatment; chi2(1)=.23, p = .630). We therefore had a final sample size of 1,030 participants
(Mage = 36), of whom 364 were male, 610 were female, 30 chose some other response (e.g.,
trans/non-binary), and 26 did not respond to the gender question.

Materials

We used 16 true and 16 false headlines to create 32 emotionally-charged, fear or anger evoking,
variants and 32 emotionally-neutral variants by manipulating the words in the headlines. We also
utilized an emotion inoculation video for the treatment condition and an educational video on
curling, bananas, or freezer burn for the control condition. Full materials can be found on OSF;
the inoculation video is available at https://inoculation.science/.

Procedure

Participants completed the first attention check in the beginning of the study and then proceeded
to provide their consent, which was followed by the second attention check question. The second
attention check question was an eliminating question and participants who failed this check were
 8

redirected to the end of the study. Participants then answered two questions on their social media
usage: the type of content they would consider sharing on social media (Political news, Sports
news, Business news, etc), and the social media platforms that they use (Facebook, Twitter,
Instagram, etc). Participants who indicated that they do not use any social media platform were
redirected to the end of study as well. Since the study entailed watching a video, we next asked
participants if the volume on their device was on. For participants who answered in the negative,
we provided an additional opportunity to turn the volume on. Those who could not or indicated
being unwilling to do so were redirected to the end of the study.

Participants were then randomly assigned to either the treatment or the control condition.
Participants in the treatment condition viewed an emotion inoculation video, whereas
participants in the control condition viewed an educational video of a similar length on one of
three topics: curling, bananas, and freezer burn. The videos were set up such that they would
autoplay and were not interactable. Therefore, participants could not pause the videos once they
automatically started playing. Moreover, the next button was not made visible till the video was
done playing.

Participants then read the instructions about viewing and rating 32 social media headlines. A
total of 64 social media headlines generated for this study were evenly split between two blocks.
The first block had 16 false and 16 true headlines, eight emotionally-charged and eight
emotionally-neutral headlines for each veracity. The second block had emotionally opposite
counterparts of the headlines in the first block. After viewing the headlines, participants rated the
trustworthiness and the manipulativeness of the headline on a 6-point scale: 1) Extremely
untrustworthy to 6) Extremely trustworthy, and 1) Not at all manipulative to 6) Extremely
manipulative.

All participants completed the third attention check question after rating the headlines.
Participants then completed two measures outside the scope of the present investigation. Namely,
the actively open-minded thinking scale (AOT, Newton et al., 2023) and the emotion regulation
questionnaire (ERQ, Spaapen et al., 2014). Participants then completed the last attention check
question, followed by questions related to their political position, and their political preference in
general, on social issues, and on economic issues. Finally, participants answered questions
related to demographics. Demographics questions entailed measurement of age, gender,
education, income, and ethnicity. Participants also indicated how much they believe in God(s).
These measures are also outside the scope of the present investigation and we do not report the
results.

We note that a first version of this experiment was run using participants recruited from Lucid
(Coppock & Mcclellan, 2019). In the Lucid experiment, participants were asked either about
manipulativeness and trustworthiness (as here), or accuracy (as in Experiment 2), or social media
sharing intentions. However, there were issues with inattentiveness (30.7% of participants who
 9

passed the initial A/V check failed at least one trivial attention check in the Lucid experiment,
compared to only 5.6% in the MTurk experiment) and significantly more participants dropped
out during the inoculation video compared to the control videos (chi2(1) = 5.26, p = 0.023)
which violates random assignment and undermines causal inference. Thus, we do not consider
the results of the Lucid experiment to provide useful insight (although for completeness, we
report them in the SI).

Our preregistration (see OSF) included political salience as a factor, but we report the analysis
without this factor included because it is not central to the research question, and including it
leads to an overly complicated model with 4-way interactions.

Results

We ran two linear regression models with robust standard errors clustered on subject and
headline, predicting (a) manipulativeness and (b) trustworthiness using an inoculation treatment
dummy (0=control, 1=treatment), a news headline veracity dummy (0=false, 1=true), an
emotional manipulativeness dummy (0=not emotional, 1=emotional), and their interactions
(Table 2). First, supporting the validity of our items, true content was judged as less manipulative
and more trustworthy than false content (b’s > 1.2, p’s < .001), and high emotion content was
judged as more manipulative and less trustworthy than low emotion content (b’s > .28, p’s <
.001) - although judgments of manipulativeness/trustworthiness were apparently much more
influenced by truth than emotionality. Second, replicating Roozenbeek et al. (2022), we also find
that the inoculation video significantly increased the difference in manipulativeness and
trustworthiness judgments for low versus high emotion content - that is, the treatment increased
emotion discernment (manipulativeness: b = 0.41, d = 0.58, p < .001; trustworthiness: b = .21, d
= 0.41, p < .001); see Figure 2. In particular, the inoculation video significantly increased
perceived manipulativeness of high emotion headlines (d = .10, p = .002) but did not
significantly decrease trustworthiness of high emotion headlines (d = -.04, p = .134). The
inoculation also significantly decreased perceived manipulativeness of low emotion headlines (d
= -.14, p < .001) and increased trustworthiness of low emotion headlines (d = .10, p < .001).

We next ask whether there was a carry-over effect of the inoculation to news headline truth
discernment; i.e., did the inoculation treatment interact significantly with veracity (either directly
or in a three-way interaction with emotional manipulativeness)? As is evident from Table 2, the
answer in both cases is no. That is, the inoculation did not interact with news headline veracity
for either manipulativeness or trustworthiness (b’s < .04, p’s > .331), nor was there a three-way
interaction between the inoculation treatment, headline veracity, and emotionality for either
measure (b’s < .05, p’s > .302). Thus, although the inoculation facilitated the detection of
emotional manipulativeness in news headlines, this effect was equivalent for true and false
headlines. Furthermore, teaching people about emotional manipulativeness does not (by itself)
have an impact on whether they can distinguish between true and false news headlines. Also, as
 10

noted, even judgments of manipulativeness were influenced more than twice as strongly by truth
than emotionality.

Table 2. OLS regression analyses (b) predicting (a) manipulativeness and (b) trustworthiness
using an inoculation treatment dummy (0=control, 1=treatment), a news headline veracity dummy
(0=false, 1=true), an emotional manipulativeness dummy (0=not emotional, 1=emotional), and
their interactions. Standard errors (clustered on participants and headlines) are in parentheses.

(a) Manipulativeness (b) Trustworthiness

(Intercept) 3.682*** 3.009***
(0.096) (0.078)
inoculation treatment -0.032 0.051
(0.051) (0.037)
headline veracity (true) -1.243*** 1.461***
(0.183) (0.151)
emotional manipulativeness 0.461*** 0.280***
(0.057) (0.052)
inoculation * veracity -0.065 0.034
(0.067) (0.050)
inoculation * emotionality 0.421*** -0.218***
(0.054) (0.026)
veracity * emotionality -0.078 0.121
(0.116) (0.106)
inoculation * veracity * 0.105 -0.045
emotionality (0.102) (0.050)

Observations 32,470 32,470

R2 0.173 0.237
Note. *p<0.05, **p<0.01, ***p<0.001
 11

Figure 2. Manipulativeness (left) and trustworthiness (right) ratings by condition and item type
in Study 1. Error bars indicate 95% confidence intervals, based on robust standard errors
clustered on subject and item.
 12

Experiment 2

Experiment 1 confirmed that the inoculation video successfully increased participants’ ability to
spot emotional manipulation in real-world true and false content. Although this effect was small
(d’s = .04-.14), the video inoculation is a relatively short intervention and it did not require any
explicit interaction from the participants.

There were, however, no effects on truth discernment per se; that is, the treatment did not
increase the ability of participants to spot misinformation. Because we manipulated truth and
emotionality orthogonally, we were able to isolate the effect of the inoculation, which appears to
teach users to identify the emotionality technique without this having an additional carry-over
effect to detecting misinformation. One issue, however, is that the identification of the
manipulativeness technique in Experiment 1 was not “spontaneous” identification in the sense
that participants were given the explicit task of searching for manipulativeness and
trustworthiness. This may have distracted users from focusing on spotting misinformation and
therefore may have blunted any carry-over effect of the inoculation on truth discernment. Thus,
in Experiment 2, we replicated the exact same design as in Experiment 1, but instead asked
participants to judge the accuracy of the claims contained in the news headlines (and, therefore,
not manipulativeness per se). This contributes valuable understanding of how inoculating against
a manipulative technique carries over to accuracy judgments, or not.

Method

Participants

Participants were again recruited from Amazon’s Mechanical Turk via Cloudresearch. In total,
2,471 participants entered the experiment and were randomly assigned to rate the accuracy of the
headlines or to indicate whether they would share them online. However, 59 participants
indicated not having social media and did not therefore qualify for the study and 10 participants
failed an attention check at the beginning of the study and did not continue. Finally, 5
participants were removed from the survey because they indicated not being able to turn on their
volume and 47 participants quit the survey prior to the intervention. A further 317 participants
dropped out and did not contribute data to the primary task. The fraction of participants dropping
out did not vary significantly across conditions (14.3% in control, 12.6% in treatment;
chi2(1)=1.50, p = .221). We therefore had a final sample size of 2,033 participants (Mage = 40),
of whom 793 were male, 1163 were female, 36 chose some other response (e.g., trans/non-
binary), and 41 did not respond to the gender question.

Materials

The materials were identical to Experiment 1.

 13

Procedure

Unlike Experiment 1, this study had a 2 (condition: treatment vs. control) x 2 (question asked:
accuracy vs. sharing) design. Instead of rating trustworthiness and manipulativeness, participants
were randomly assigned to either rate the accuracy of the headline (using a scale of 1 =
extremely inaccurate to 6 = extremely accurate) or indicate how likely they would be to share the
headline on social media (using a scale of 1 = extremely unlikely to 6 = extremely likely). The
procedure of this study was identical to Experiment 1 in all other aspects.

We focus here on the results for the accuracy condition. The results for the sharing condition are
presented in the SI. In short, the inoculation did not increase either emotion or truth discernment
for sharing 2.

Results

We ran a linear regression model with robust standard errors clustered on subject and headline,
predicting accuracy judgments using an inoculation treatment dummy (0=control, 1=treatment),
a news headline veracity dummy (0=false, 1=true), an emotional manipulativeness dummy
(0=not emotional, 1=emotional), and their interactions (Table 3). Similar to Experiment 1, true
content was judged as more accurate than false content (b = 0.38, p < .001), and high emotion
content was judged as less accurate than low emotion content (b = .05, p < .001). Second, the
inoculation video significantly increased the difference in perceived accuracy for low versus high
emotion content - that is, the treatment increased emotion discernment (b = -.02, p = .001).
However, as in Experiment 1, this effect was fairly small (see Figure 3). In particular, the
inoculation video significantly decreased perceived accuracy of high emotion headlines (d = -.04,
p = .049) but had no significant effect on perceived accuracy of low emotion headlines (d = .02,
p = .345).

2 Note that the preregistration for Experiment 1 mentions the Experiment 2 dependent variables. We ran Experiment
1 twelve days prior to Experiment 2 (to ensure successful technique recognition), but preregistered them together as
one Experiment with 3 dependent variables. We report them separately here for simplicity.
 14

Table 3. OLS regression analysis (b) predicting accuracy judgments using an inoculation
treatment dummy (0=control, 1=treatment), a news headline veracity dummy (0=false, 1=true),
an emotional manipulativeness dummy (0=not emotional, 1=emotional), and their interactions.
Standard errors (clustered on participants and headlines) are in parentheses.

Perceived accuracy
(Intercept) 0.437***
(0.016)
inoculation treatment -0.004
(0.006)
headline veracity (true) 0.380***
(0.032)
emotional manipulativeness -0.049***
(0.010)
inoculation * veracity -0.008
(0.010)
inoculation * emotionality -0.019**
(0.006)
veracity * emotionality 0.043*
(0.021)
inoculation * veracity * -0.001
emotionality (0.012)

Observations 32,058
R2 0.315
Note. *p<0.05, **p<0.01, ***p<0.001
 15

Figure 3. Perceived accuracy by condition and item type in Study 2. Error bars indicate 95%
confidence intervals, based on robust standard errors clustered on subject and item.

The inoculation treatment did not, however, have a carry-over effect on truth discernment.
Specifically, there was no interaction between the inoculation and news headline veracity (b = -
.01, p = .446), nor was there a three-way interaction between the inoculation treatment, headline
veracity, and emotionality (b’s = -.001, p = .919).

As a follow-up, we also restricted the analysis to only false-high emotion and true-low emotion
items; this represents a maximally favorable limiting case where emotional manipulation is
always present among false content and never present among true content. Thus, in this case,
decreasing belief in emotional content should be expected to also decrease belief in false content
(and therefore increase discernment). However, even here, we still did not find a significant
effect of the inoculation on truth discernment (interaction between condition and veracity, b =
.011, p = .406). Accordingly, there was no significant inoculation effect on accuracy judgments
for either false-high emotion (d = -0.03, p = .401) or true-low emotion (d = 0.01, p = .815)
headlines.
 16

Experiment 3

Experiment 2 confirmed that the inoculation video decreased belief in emotionally manipulative
content relative to non-emotionally manipulative content. Thus, the effects of Experiment 1 (and
that of Roozenbeek et al., 2022) do not appear to be driven entirely by explicitly focusing
participants on the task of identifying manipulativeness (by asking about it directly instead of
asking about accuracy, as we did here) 3. However, as in Experiment 1, there again were no
effects on truth discernment. Surprisingly, there was no effect even when only analyzing the
subset of items where all false headlines were emotionally manipulative and all true headlines
were not. Thus, even if one were to assume that the posited misinformation techniques are only
present for misinformation (and therefore presenting the most favorable context for testing the
effect of the intervention), the inoculation did not apparently help participants spot
misinformation.

To test this more directly, we re-ran the experiment but only including the high emotionality
false and low emotionality true headlines. This was intended to rule out the possibility that
including the fully crossed set of true/false and high/low emotionality headlines undermined the
effect of the intervention in some way 4.

Method

Participants

U.S. participants were recruited from Prolific Academic. In total, 1,526 participants entered the
experiment and were randomly assigned to rate the accuracy of the headlines or to indicate
whether they would share them online. However, 48 participants indicated not having social
media and did not therefore qualify for the study and 7 participants failed an attention check at
the beginning of the study and did not continue. Finally, no participants were removed from the
survey because they indicated not being able to turn on their volume, but 9 participants quit the
survey prior to the intervention. A further 254 participants dropped out and did not contribute
data to the primary task. The fraction of participants dropping out did not vary significantly
across conditions (17.2% in control, 17.6% in treatment; chi2(1)=0.04, p = .841). We therefore
had a final sample size of 1,208 participants (Mage = 40), of whom 652 were male, 517 were
female, 28 chose some other response (e.g., trans/non-binary), and 11 did not respond to the
gender question.

3 Of course, participants were nonetheless asked directly about accuracy – see SI for evidence that the inoculation
videos did not have the intended effect when participants were asked about sharing the content online.
4 We ran a parallel experiment using participants from Mechanical Turk (N = 1,023) that is not reported in the SI.

We similarly did not find an effect of the emotional inoculation (b = -0.003, p = .734); however, we believed that the
data were questionable because the difference between true and false headlines, which is robustly observed to be
large in all other studies in this paper as well as in the literature more generally, was extremely small (b = .04,
relative to b = .48 in the parallel Experiment 3). Thus we do not consider this study further.
 17

Materials and Procedure

Out of the 32 emotionally-charged (16 true and 16 false) and 32 emotionally-neutral (16 true and
16 false) headlines used in Experiment 1, we kept only the 16 false emotionally-charged
headlines and the 16 true emotionally-neutral headlines for use in this study.

Participants provided accuracy ratings as in Experiment 2 (there was no sharing condition). The
16 headlines seen by the participants came from one of two randomly-chosen blocks. The first
block had eight false emotionally-charged and eight true emotionally-neutral headlines. The
second block had a different set of false emotionally-charged and true emotionally-neutral
headlines, eight of each kind. The remaining procedure was identical to Experiment 2, except
that we removed the AOT, ERQ, and the fourth attention check (owing to the shorter length).

The analyses were preregistered (see OSF).

Results

In this experiment, emotional manipulativeness and headline truth were manipulated to be fully
consistent with each other. Thus, we ran a linear regression model with robust standard errors
clustered on subject and headline, predicting accuracy judgments using an inoculation treatment
dummy (0=control, 1=treatment), a news headline veracity/emotional manipulativeness dummy
(0=false/high emotional, 1=true/low emotional), and their interaction (Table 4). As expected,
true low emotional content was judged as more accurate than false high emotional content (b =
0.48, p < .001). However, the inoculation video did not significantly increase the difference in
perceived accuracy for true low emotion versus false high emotion content (b = -0.01, p = .559).
That is, participants did not improve in their ability to distinguish between true and false news
headlines even though we manipulated the true headlines to be low emotion and the false
headlines to be high emotion. There was no significant inoculation effect on accuracy judgments
for either false-high emotion (d = -0.01, p = .691) or true-low emotion (d = 0.01, p = .709)
headlines (see Figure 4).
 18

Table 4. OLS regression analysis (b) predicting accuracy judgments using an inoculation
treatment dummy (0=control, 1=treatment), a news headline veracity/emotional
manipulativeness dummy (0=false/high emotional, 1=true/low emotional), and their interaction.
Standard errors (clustered on participants and headlines) are in parentheses.

Perceived accuracy
(Intercept) 0.163***
(0.024)
inoculation treatment 0.003
(0.009)
headline manipulativeness 0.479***
(0.031)
inoculation * manipulativenss -0.007
(0.012)

Observations 19,266
R2 0.446
Note. *p<0.05, **p<0.01, ***p<0.001

Figure 4. Perceived accuracy by condition and item type in Study 3. Error bars indicate 95%
confidence intervals, based on robust standard errors clustered on subject and item.
 19

Experiment 4

Experiment 3 set a highly favorable context for testing the effect of the inoculation on discerning
between true and false news headlines – i.e., participants only received false headlines that
contained the emotional manipulation technique and true headlines that did not contain the
technique; hence, identifying the technique should correspond directly to detecting
misinformation. Notwithstanding, the inoculation did not produce a significant difference
between true (low emotion) and false (high emotion) videos. Thus, the emotional language
inoculation, at least in the present form, appears to be very specific: It helps people identify
emotionality in a general sense (as evidenced in Experiments 1 and 2), but its effect is
undermined by some combination of (a) asking about accuracy (instead of the presence of the
technique per se) and (b) the presence of claims that are either true or false (rather than the
techniques presented on their own without claims of fact). That is, when the task is made more
difficult by intermixing actual true or false claims, the video appears to lose its effectiveness as
an “inoculation against misinformation”.

One potential reason for this is that the emotional language inoculation video does not actually
draw attention to accuracy in any specific way. Rather, the video focuses on identifying
emotionality techniques as a form of grabbing one’s attention. Although the videos were
proposed as a misinformation intervention, it may be that they do not actually draw people’s
attention to whether content is true or false. Other research has shown that, at least in terms of
decisions about what to share online, users may often fail to even consider whether content is
accurate before they decide to share that content (Pennycook et al., 2021). Indeed, subtly
reminding people about accuracy is sufficient to increase the quality of content (i.e., less false
relative to true content) that people are willing to share online (Pennycook & Rand, 2022). One
possibility, then, is that a simple reminder about accuracy (and the threat of misinformation)
prior to the inoculation video may help users apply the specific content that they learn about
emotional manipulation in the video. In support of this possibility, previous inoculation research
that used longer videos with a broader range of techniques and that also included explicit
reference to accuracy has found an effect on judgments of the reliability (a proxy for accuracy)
of subsequent misinforming material (Lewandowsky & Yesilada, 2021) (albeit, as discussed,
material that does not contain true or false claims per se). To test this, we modified the
inoculation video so that it was bookended with reminders about accuracy (video can be viewed
at OSF).

Method

Participants

U.S. participants were recruited from Prolific Academic. In total, 1,348 participants entered the
experiment and were randomly assigned to rate the accuracy of the headlines or to indicate
whether they would share them online. However, 27 participants indicated not having social
 20

media and did not therefore qualify for the study and 4 participants failed an attention check at
the beginning of the study and did not continue. Finally, no participants were removed from the
survey because of the intervention. A further 95 participants dropped out and did not contribute
data to the primary task. The fraction of participants dropping out did not vary significantly
across conditions (6.8% in control, 7.8% in treatment; chi2(1)=0.45, p = .504). We therefore had
a final sample size of 1,211 participants (Mage = 40), of whom 647 were male, 514 were female,
37 chose some other response (e.g., trans/non-binary), and 13 did not respond to the gender
question.

Materials and Procedure

Videos used in this study were modified to include accuracy prompts. This new video had
accuracy reminders in the beginning and at the end, with the inoculation video used in the
previous experiments ‘sandwiched’ in the middle (see OSF). The three educational videos in the
control condition were also replaced to match the duration of the new treatment video. The new
educational control videos were about stars, Scotland, and bats. Headlines and procedure were
the same as in Experiment 3. The analyses were preregistered (see OSF).

Results

As in Experiment 3, emotional manipulativeness and headline truth were manipulated to be fully

consistent with each other. Thus, we ran a linear regression model with robust standard errors
clustered on subject and headline, predicting accuracy judgments using an inoculation (with
accuracy prompt) treatment dummy (0=control, 1=treatment), a news headline
veracity/emotional manipulativeness dummy (0=false/high emotional, 1=true/low emotional),
and their interaction (Table 5). True low emotional content was judged as more accurate than
false high emotional content (b = 0.45, p < .001). Critically, the inoculation video did
significantly increase the difference in perceived accuracy for true low emotion versus false high
emotion content (b = 0.03, d = 0.16, p = .003). That is, the combined inoculation and accuracy
prompt video was successful in leading participants to improve in their ability to distinguish
between true and false news headlines (when true headlines were manipulated to be low emotion
and false headlines were manipulated to be high emotion). Interestingly, the inoculation had no
effect on accuracy judgments for false-high emotion headlines (d = -0.036, p = .282) but
significantly increased accuracy judgments for true-low emotion headlines (d = 0.089, p = .005)
(see Figure 5).
 21

Table 5. OLS regression analyses (b) predicting accuracy judgments using a combined
inoculation and accuracy prompt treatment dummy (0=control, 1=treatment), a news headline
veracity/emotional manipulativeness dummy (0=false/high emotional, 1=true/low emotional),
and their interaction. Standard errors (clustered on participants and headlines) are in
parentheses.

Perceived accuracy
(Intercept) 0.169***
(0.022)
inoculation & accuracy prompt treatment -0.009
(0.008)
headline manipulativeness 0.448***
(0.030)
inoculation & accuracy prompt * manipulativeness 0.034**
(0.011)

Observations 19,336
R2 0.438
Note. *p<0.05, **p<0.01, ***p<0.001

Figure 5. Perceived accuracy by condition and item type in Study 4. Error bars indicate 95%
confidence intervals, based on robust standard errors clustered on subject and item.
 22

Experiment 5

Experiment 4 provided evidence that the inoculation was effective at increasing truth
discernment when paired with an accuracy prompt. Although this effect was small and evident
only for true-low emotion headlines (as opposed to false-high emotion headlines), it nonetheless
provides some evidence that participants were able to apply what they learned about emotional
manipulation in the short video to content that contains true and false claims. Nonetheless, one
question remains: Was the accuracy prompt itself responsible for the increase in truth
discernment, or are both elements necessary to produce the effect? We test this in Experiment 5
by comparing the combined accuracy prompt and inoculation video with the accuracy prompt
video on its own.

Method

Participants

U.S. participants were recruited from Prolific Academic. In total, 1,996 participants entered the
experiment and were randomly assigned to rate the accuracy of the headlines or to indicate
whether they would share them online. However, 62 participants indicated not having social
media and did not therefore qualify for the study and 4 participants failed an attention check at
the beginning of the study and did not continue. Finally, 1 participant was removed from the
survey because they indicated not being able to turn on their volume, and 11 participants quit the
survey prior to the volume question. A further 114 participants dropped out and did not
contribute data to the primary task. The fraction of participants dropping out did not vary
significantly across conditions (6.0% in long control, 6.6% in combined inoculation and
accuracy, 5.5% in short control, 5.8% in accuracy only; chi2(3)=0.59, p = .899). We therefore
had a final sample size of 1,804 participants (Mage = 40), of whom 907 were male, 838 were
female, 49 chose some other response (e.g., trans/non-binary), and 11 did not respond to the
gender question.

Materials and Procedure.

In this study, we had two sets of videos. As in Experiment 4, we had a ‘sandwich’ inoculation
video with accuracy reminders in the beginning and at the end for the treatment condition, and
three duration-matched educational videos (on stars, Scotland, and bats) for the control
condition. Unlike Experiment 4, we additionally created a condition where participants only
received the initial accuracy-prompt component of the video. Consequently, we used a different
set of three duration-matched educational/motivational videos (on recycling, entrepreneurship,
and effort) for the control condition. The materials and procedure were otherwise identical to
Experiment 4. The analyses were preregistered (see OSF).
 23

Results

As preregistered, we first confirmed that there were no differences between the two sets of
control videos (condition dummy: b = =0.051, p = 0.366; interaction between condition and
veracity: b = 0.0166, p = 0.807). Thus, we pooled the control conditions. We ran a linear
regression model with robust standard errors clustered on subject and headline, predicting
accuracy judgments using a combined inoculation and accuracy prompt treatment dummy
(0=control, 1=treatment), an only accuracy prompt treatment dummy (0=control, 1=treatment), a
news headline veracity/emotional manipulativeness dummy (0=false/high emotional, 1=true/low
emotional), and the interactions between manipulativeness and our two treatment conditions
(Table 6). True low emotional content was judged as more accurate than false high emotional
content (b = 0.45, p < .001).

Critically, replicating Experiment 4, the inoculation video significantly increased the difference
in perceived accuracy for true low emotion versus false high emotion content (b = 0.05, d = 0.22,
p = .001). In this case, the combined inoculation and accuracy prompt video had (marginally)
significantly decreased perceptions of accuracy for false-high emotion headlines (d = -0.088, p =
.060) and significantly increased accuracy judgments for true-low emotion headlines (d = 0.090,
p = .033) (see Figure 6). As expected, there was no interaction effect for the accuracy prompt
alone (b = 0.001, p = .948), indicating that simply drawing people’s attention to the problem of
misinformation is not sufficient to increase subsequent truth discernment. Importantly, the effect
of the combined treatment on discernment was significantly larger than the effect of the accuracy
prompt alone (Wald test comparing the coefficient on (i) the interaction between accuracy
prompt only and veracity with (ii) the interaction between combined inoculation+accuracy and
veracity, chi2(1) = 11.33, p = .0008) In other words, the accuracy prompt and inoculation
treatment had a synergistic effect on truth discernment, but similar to prior experiments, neither
were effective on their own. Interestingly, since prior work on accuracy prompts has focused on
news headline sharing (and not belief, as we measure here), this research also indicates that
inoculations may boost the effect of prompting accuracy beyond choices about what to share.
 24

Table 6. OLS regression analyses (b) predicting accuracy judgments using an inoculation
treatment dummy (0=control, 1=treatment), a news headline veracity/emotional
manipulativeness dummy (0=false/high emotional, 1=true/low emotional), and their interaction.
Standard errors (clustered on participants and headlines) are in parentheses.

Perceived accuracy
(Intercept) 0.175***
(0.025)
inoculation & accuracy prompt treatment -0.022*
(0.011)
accuracy prompt treatment -0.009
(0.010)
headline manipulativeness 0.452***
(0.031)
inoculation & accuracy prompt * manipulativeness 0.046**
(0.013)
accuracy prompt * manipulativeness 0.001
(0.011)

Observations 28,8812
R2 0.438
Note. *p<0.05, **p<0.01, ***p<0.001
 25

Figure 6. Perceived accuracy by condition and item type in Study 5. Error bars indicate 95%
confidence intervals, based on robust standard errors clustered on subject and item.

General Discussion

A common sentiment is that multiple tools will be needed to effectively combat misinformation
(Bak-Coleman et al., 2022; Bode & Vraga, 2021). We present evidence here that popular
intervention approaches can have synergistic effects, even when not effective on their own.

In particular, we tested a popular inoculation intervention intended to increase people’s ability to

recognize emotional language manipulation in text. Unlike past work, we tested this intervention
using stimuli that not only contained the emotional language manipulation (in contrast to
neutral/low emotionality content), but that also contained true or false claims - allowing us to
assess the effect of the inoculation on truth discernment.

Our results identified a limitation of the inoculation intervention. Consistent with past work (e.g.
Roozenbeek et al., 2022), the intervention was successful in helping people distinguish between
content that contains (vs. does not contain) an emotional language manipulation. Critically,
however, we found consistent evidence that the inoculation on its own did not help people
distinguish between true and false content. In fact, we did not find an inoculation effect on
accuracy discernment even in the extreme case where all false claims - and no true claims -
contained emotional manipulation.
 26

Why did the inoculation fail to facilitate truth discernment? One plausible mechanism is that the
emotional inoculation that we focused on (which was among the strongest interventions in past
work; Roozenbeek et al., 2022) did not draw an explicit connection between the use of emotional
language and falsehoods per se. This is important because past work shows that people tend to
be inattentive to accuracy on social media (Epstein et al., 2023). Consistent with this, we found
that focusing people’s attention on accuracy prior to watching the inoculation video (as is done
in popular accuracy prompt interventions; Pennycook & Rand, 2022) did lead to an increase in
truth discernment - as a consequence of the combined force of inoculation and accuracy
prompting. This demonstrates a synergy between inoculation and accuracy prompt interventions.

Implications

These findings emphasize the importance of jointly deploying multiple interventions. Building
off recent theorizing about boost versus nudge interventions, our results suggest that an
inoculation “boost” - i.e., an intervention that increases people’s cognitive and motivational
competencies through learning (Hertwig & Grüne-Yanoff, 2017) - is important for improving
truth discernment capabilities, whereas an accuracy prompt nudge is critical for
engaging/activating that capability.

Our findings have implications for work on psychological inoculation in the context of
misinformation. In particular, misinformation inoculation interventions have typically been
evaluated by testing their effects on people’s ability to identify manipulative techniques. This is
usually done using stylized or “synthetic” materials that do not contain actual true or false
claims, in an effort to offer “clean” examples of the manipulative technique absent real-world
truth claims that might influence people’s evaluations.

The logic behind this approach is that helping people recognize common manipulation
techniques, such as emotional language manipulation, ought to help them distinguish a wide
range of misinformation, which includes misleading content as well as false content. Our
findings, however, indicate that simply providing information about manipulative techniques
may not be sufficient to lead people to be “inoculated” against belief in actual falsehoods.
Rather, only when people were prompted to connect the information about emotional
manipulation with factuality per se did we find evidence that the inoculation had an impact on
truth discernment.

The implication of this is that prior inoculation approaches may only be broadly effective if they
have successfully connected the information that people are learning about manipulation
techniques with the threat of misinformation. Moreover, even in such cases, past work that has
tested inoculation approaches using technique identification may have created inflated
expectations regarding effect sizes. For example, Experiment 1 found effect sizes of d = 0.41
and d = 0.58 on technique discernment using manipulativeness and trustworthiness ratings as the
outcome (in line with Roozenbeek et al.’s (2022) finding of d = 0.49 on technique discernment).
 27

In contrast, however, Experiment 2 found an effect size of only d = 0.20 on technique

discernment when using identical materials but with perceived accuracy as the outcome - and, as
mentioned above, found no significant effect (d = -0.03) on truth discernment. Furthermore,
when we did ultimately find a significant improvement in truth discernment from the combined
inoculation+accuracy prompt intervention in Studies 4 and 5, this effect was driven by increased
belief in low-emotion true headlines more so than decreased belief in high-emotion false
headlines. This may be an instance of the well-known “mirror effect” in recognition memory
(Glanzer & Adams, 1985, 1990), which occurs when manipulations of memory strength affect
not only the memorized material, but also new items that are interleaved with the memoranda at
test and serve as foils. Nonetheless, future work should reconcile this pattern with inoculation
theory’s focus on reducing susceptibility to manipulative content per se.

Limitations

One important empirical limitation of our research is that we found a successful inoculation
(with accuracy prompt) effect on truth discernment in a context where the emotional language
manipulation was fully confounded with truth. That is, the false content all contained an
emotional language manipulation, and the true content did not contain any emotional language
manipulation. There is reason to expect false content to be somewhat more likely to contain
emotional manipulation than true content. For example, Carrasco-Farré (2022) showed in
analysis of nearly 100,000 news articles that fake news articles contain more negative sentiment
on average than factual articles. However, emotional language and truth/falsity are certainly far
from perfectly correlated in the real world. Thus, to whatever extent a disconnect exists between
emotionality and veracity outside the lab setting, the true effect on truth discernment will be
smaller than the average effect of d = 0.20 observed for the combined intervention in our studies.

Another limitation of the present studies is the reliance on headlines alone, which provides a very
limited attack surface for the inoculation intervention to find traction. It is therefore possible that
a more realistic extent of content (e.g., a headline plus a lede or even an entire news article) will
provide more opportunity for the intervention to “grab”. Although this remains an open empirical
question, it is quite plausible that inoculation effect sizes may be greater when there is more
content that the intervention can be applied to.

Furthermore, this study only tested one medium and format of interventions for both inoculation
and accuracy prompts - animated video. It is worth testing variations of videos - live action,
creator-driven, humorous - as well as other mediums entirely such as text, audio, or captioned
images. Similarly, we used a specific set of 32 headlines. Future work should investigate how the
results generalize to other headline sets.

Future work should look into the mechanism at play in the accuracy prompt used in our study. It
is unclear whether the prompt in our study was priming accuracy (as in past work on sharing
intentions) or something else, such as linking the inoculation to misinformation. The portion of
 28

the video including the prime did mention accuracy, but accuracy was not the primary focus.
Past work has used a variety of accuracy prompt methods (e.g., Epstein et al., 2021; Pennycook
& Rand, 2022); however, one limitation is that these tend to be varied and may differ in terms of
how closely connected they are to the key mechanism(s) of interest. Future work clarifying this
issue would be beneficial.

Conclusion

We conclude that combining interventions can have synergies for building resilience to
misinformation. Theoretical modeling has suggested the existence of synergies from combining
approaches to fight misinformation (Bak-Coleman et al., 2022; Bode & Vraga, 2021). The
present studies present empirical evidence of this phenomenon using randomized-controlled
trials with participants viewing scalable interventions and real world headlines. Our study has
significant implications for the burgeoning field designing misinformation interventions. The
practitioners, technologists, and policymakers developing these interventions would benefit from
testing and deploying multiple methods to fight misinformation in tandem where they are
synergistic. Combined interventions can remain highly scalable; in this case, both interventions
were delivered together in a short video of about two minutes.

Notwithstanding the potential success of such scalable interventions, they can only be part of the
solution to the misinformation problem. Misinformation does not arise in a vacuum but
flourishes as part of an online ecosystem whose architecture focuses on capturing user attention
(e.g., Kozyreva et al., 2020). Given that purveyors of misinformation are, by definition, freed
from constraints imposed by reality, they can exploit known biases in human attention (e.g., by
presenting outraging content (e.g., Bakir & McStay, 2018), which platform algorithms are likely
to amplify). Interventions such as those developed here must therefore be accompanied by
creating an internet with democratic credentials and based on user empowerment (Lewandowsky
& Pomerantsev, 2022).

Acknowledgements

This research was supported by funding from Jigsaw (Google). GP acknowledges financial
support from the Social Sciences and Humanities Research Council of Canada, the John
Templeton Foundation and the Office of Naval Research. DR and AB acknowledge support from
the National Science Foundation (NSF Award 2047152) and the Alfred P. Sloan Foundation
(Grant #2021-16891). SL acknowledges financial support from the European Research Council
(ERC Advanced Grant 101020961 PRODEMINFO), the Humboldt Foundation through a
research award, the Volkswagen Foundation (grant ``Reclaiming individual autonomy and
democratic discourse online: How to rebalance human and algorithmic decision making''), and
the European Commission (Horizon 2020 grant 101094752 SoMe4Dem). SL also receives
funding from UK Research and Innovation (through EU Horizon replacement funding grant
number 10049415). The views and conclusions contained herein are those of the authors and
 29

should not be interpreted as representing the official policies, either expressed or implied, of the
funding bodies.

Competing interests

GP and DR have received funding from Meta and Google; GP, AB, and RC were employed at
Google during the time that the studies were designed and the data was collected, and BG both
was and is employed at Google. Google has invested in both inoculation and accuracy prompts
as misinformation interventions.

References

Arechar, A. A., Allen, J. N. L., Berinsky, A., Cole, R., Epstein, Z., Garimella, K., Gully, A., Lu,
J. G., Ross, R. M., Stagnaro, M. N., Zhang, J., Pennycook, G., & Rand, D. G. (2023).
Understanding and combating online misinformation across 16 countries on six
continents. Nature Human Behaviour, 1–48. https://doi.org/10.31234/OSF.IO/A9FRZ

Athey, S., Grabarz, K., Luca, M., & Wernerfelt, N. (2023). Digital public health interventions at
scale: The impact of social media advertising on beliefs and outcomes related to COVID
vaccines. Proceedings of the National Academy of Sciences, 120(5), e2208110120.
https://doi.org/10.1073/pnas.2208110120

Bak-Coleman, J. B., Kennedy, I., Wack, M., Beers, A., Schafer, J. S., Spiro, E. S., Starbird, K.,
& West, J. D. (2022). Combining interventions to reduce the spread of viral
misinformation. Nature Human Behaviour, 6(10), Article 10.
https://doi.org/10.1038/s41562-022-01388-6

Bakir, V., & McStay, A. (2018). Fake News and The Economy of Emotions. Digital Journalism,
6(2), 154–175. https://doi.org/10.1080/21670811.2017.1345645

Bode, L., & Vraga, E. (2021). The Swiss cheese model for mitigating online misinformation.
Bulletin of the Atomic Scientists, 77(3), 129–133.
https://doi.org/10.1080/00963402.2021.1912170

Chan, M. pui S., Jones, C. R., Hall Jamieson, K., & Albarracín, D. (2017). Debunking: A Meta-
Analysis of the Psychological Efficacy of Messages Countering Misinformation.
Psychological Science, 28(11), 1531–1546. https://doi.org/10.1177/0956797617714579

Compton, J. (2013). Inoculation Theory. In J. P. Dillard & L. Shen (Eds.), The SAGE Handbook
of Persuasion: Developments in Theory and Practice (pp. 220–236).
https://books.google.ca/books?hl=en&lr=&id=ZO1yAwAAQBAJ&oi=fnd&pg=PT226&
dq=+Inoculation+theory&ots=-iu3ZmG7m5&sig=XYVk-
XasWmjjVFcQJgEfvqxDKIY#v=onepage&q=Inoculation theory&f=false
 30

Compton, J., van der Linden, S., Cook, J., & Basol, M. (2021). Inoculation theory in the post-
truth era: Extant findings and new frontiers for contested science, misinformation, and
conspiracy theories. Social and Personality Psychology Compass, 15(6), e12602.
https://doi.org/10.1111/spc3.12602

Cook, J., Lewandowsky, S., Ecker, U., Jacobs, P., Cowtan, K., & Green, S. (2017). Neutralizing
misinformation through inoculation: Exposing misleading argumentation techniques
reduces their influence. PloS One, 12(5), e0175799.
https://doi.org/10.1371/journal.pone.0175799

Coppock, A., & Mcclellan, O. A. (2019). Validating the Demographic, Political, Psychological,
and Experimental Results Obtained from a New Source of Online Survey Respondents.
Research & Politics.

Douglas, B. D., Ewell, P. J., & Brauer, M. (2023). Data quality in online human-subjects
research: Comparisons between MTurk, Prolific, CloudResearch, Qualtrics, and SONA.
PLOS ONE. https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0279720

Ecker, U. K. H., Lewandowsky, S., Cook, J., Schmid, P., Fazio, L. K., Brashier, N., Kendeou, P.,
Vraga, E. K., & Amazeen, M. A. (2022). The psychological drivers of misinformation
belief and its resistance to correction. Nature Reviews Psychology, 1(1), Article 1.
https://doi.org/10.1038/s44159-021-00006-y

Epstein, Z., Berinsky, A. J., Cole, R., Gully, A., Pennycook, G., & Rand, D. G. (2021).
Developing an accuracy-prompt toolkit to reduce COVID-19 misinformation online.
Harvard Kennedy School Misinformation Review. https://doi.org/10.37016/mr-2020-71

Epstein, Z., Sirlin, N., Arechar, A., Pennycook, G., & Rand, D. (2023). The social media context
interferes with truth discernment. Science Advances, 9(9), eabo6169.
https://doi.org/10.1126/sciadv.abo6169

Glanzer, M., & Adams, J. K. (1985). The mirror effect in recognition memory. Memory &
Cognition, 13(1), 8–20. https://doi.org/10.3758/BF03198438

Glanzer, M., & Adams, J. K. (1990). The mirror effect in recognition memory: Data and theory.
Journal of Experimental Psychology: Learning, Memory, and Cognition, 16(1), 5–16.
https://doi.org/10.1037/0278-7393.16.1.5

Hertwig, R., & Grüne-Yanoff, T. (2017). Nudging and Boosting: Steering or Empowering Good
Decisions. Perspectives on Psychological Science, 12(6), 973–986.
https://doi.org/10.1177/1745691617702496

Jigsaw. (2023, February 14). Defanging Disinformation’s Threat to Ukrainian Refugees. Jigsaw.
https://medium.com/jigsaw/defanging-disinformations-threat-to-ukrainian-refugees-
 31

b164dbbc1c60

Klepper, D. (2023, February 13). Google to expand misinformation ‘prebunking’ in Europe. AP

News. https://apnews.com/article/technology-science-politics-germany-business-
a10273eea5a0227c38187cc4f84d8788

Kozyreva, A., Lewandowsky, S., & Hertwig, R. (2020). Citizens Versus the Internet:
Confronting Digital Challenges With Cognitive Tools. Psychological Science in the
Public Interest, 21, 103–156. https://doi.org/10.1177/1529100620946707

Kozyreva, A., Lorenz-Spreen, P., Herzog, S., Ecker, U., Lewandowsky, S., & Hertwig, R.
(2022). Toolbox of Interventions Against Online Misinformation and Manipulation.
PsyArXiv. https://doi.org/10.31234/osf.io/x8ejt

Lewandowsky, S., Ecker, U. K. H., Seifert, C. M., Schwarz, N., & Cook, J. (2012).
Misinformation and its correction: Continued influence and successful debiasing.
Psychological Science in the Public Interest, 13(3), 106–131.
https://doi.org/10.1177/1529100612451018

Lewandowsky, S., & Pomerantsev, P. (2022). Technology and democracy: A paradox wrapped
in a contradiction inside an irony. Memory, Mind & Media, 1, e5.
https://doi.org/10.1017/mem.2021.7

Lewandowsky, S., & van der Linden, S. (2021). Countering Misinformation and Fake News
Through Inoculation and Prebunking. European Review of Social Psychology, 32(2),
348–384. https://doi.org/10.1080/10463283.2021.1876983

Lewandowsky, S., & Yesilada, M. (2021). Inoculating against the spread of Islamophobic and
radical-Islamist disinformation. Cognitive Research: Principles and Implications, 6(1),
57. https://doi.org/10.1186/s41235-021-00323-z

Lin, H., Pennycook, G., & Rand, D. G. (2023). Thinking more or thinking differently? Using
drift-diffusion modeling to illuminate why accuracy prompts decrease misinformation
sharing. Cognition, 230, 105312. https://doi.org/10.1016/j.cognition.2022.105312

Mukherjee, S., Coulter, M., & Mukherjee, S. (2022, December 6). Exclusive: Google launches
anti-misinformation campaign in India. Reuters.
https://www.reuters.com/technology/google-launches-anti-misinformation-campaign-
india-2022-12-06/

Newton, C., Feeney, J., & Pennycook, G. (2023). On the disposition to think analytically: Four
distinct intuitive-analytic thinking styles. Personality & Social Psychology Bulletin, 1–
141. https://doi.org/10.31234/OSF.IO/R5WEZ

Nieminen, S., & Rapeli, L. (2019). Fighting Misperceptions and Doubting Journalists’
 32

Objectivity: A Review of Fact-checking Literature. Political Studies Review, 17(3), 296–

309. https://doi.org/10.1177/1478929918786852

Pennycook, G., Epstein, Z., Mosleh, M., Arechar, A. A., Eckles, D., & Rand, D. (2021). Shifting
attention to accuracy can reduce misinformation online. Nature, 1–63.
https://doi.org/10.1038/s41586-021-03344-2

Pennycook, G., McPhetres, J., Zhang, Y., Lu, J. G., & Rand, D. G. (2020). Fighting COVID-19
misinformation on social media: Experimental evidence for a scalable accuracy nudge
intervention. Psychological Science. https://doi.org/10.31234/OSF.IO/UHBK9

Pennycook, G., & Rand, D. G. (2021). The Psychology of Fake News. Trends in Cognitive
Sciences, 1–29. https://doi.org/10.1016/j.tics.2021.02.007

Pennycook, G., & Rand, D. G. (2022). Accuracy prompts are a replicable and generalizable
approach for reducing the spread of misinformation. Nature Communications, 13(1), 1–
12. https://doi.org/10.1038/s41467-022-30073-5

Porter, E., & Wood, T. J. (2021). The global effectiveness of fact-checking: Evidence from
simultaneous experiments in Argentina, Nigeria, South Africa, and the United Kingdom.
Proceedings of the National Academy of Sciences of the United States of America,
118(37). https://doi.org/10.1073/PNAS.2104235118/-/DCSUPPLEMENTAL

Roozenbeek, J., Freeman, A. F., & van der Linden, S. (2021). How accurate are accuracy nudge
interventions? A pre-registered direct replication of Pennycook et al. (2020).
Psychological Science.

Roozenbeek, J., & van der Linden, S. (2022). How to Combat Health Misinformation: A
Psychological Approach. American Journal of Health Promotion, 36(3), 569–575.
https://doi.org/10.1177/08901171211070958

Roozenbeek, J., van der Linden, S., Goldberg, B., Rathje, S., & Lewandowsky, S. (2022).
Psychological inoculation improves resilience against misinformation on social media.
Science Advances, 8(34), eabo6254. https://doi.org/10.1126/sciadv.abo6254

Spaapen, D. L., Waters, F., Brummer, L., Stopa, L., & Bucks, R. S. (2014). The Emotion
Regulation Questionnaire: Validation of the ERQ-9 in two community samples.
Psychological Assessment, 26(1), 46–54. https://doi.org/10.1037/a0034474

Traberg, C. S., Harjani, T., Basol, M., Biddlestone, M., Maertens, R., Roozenbeek, J., & van der
Linden, S. (2023). Prebunking Against Misinformation in the Modern Digital Age. In T.
D. Purnat, T. Nguyen, & S. Briand (Eds.), Managing Infodemics in the 21st Century:
Addressing New Public Health Challenges in the Information Ecosystem (pp. 99–111).
Springer International Publishing. https://doi.org/10.1007/978-3-031-27789-4_8
 33

Traberg, C. S., Roozenbeek, J., & van der Linden, S. (2022). Psychological Inoculation against
Misinformation: Current Evidence and Future Directions. The Annals of the American
Academy of Political and Social Science.

Wood, T., & Porter, E. (2019). The elusive backfire effect: Mass attitudes’ steadfast factual
adherence. Political Behavior, 41(1), 135–163.
 Experiment 1 & 2 with sample from Lucid

Method

Methods were identical to Experiment 1 and 2 (as reported in the main text). Participants (N = 1,922 from
Lucid) were randomized to either rate perceived manipulativeness/trustworthiness, perceived accuracy, or
sharing intentions, and to receive either the inoculation video or a control video.

Results
There was no effect of the inoculation on accuracy judgments, and no significant interactions.
For sharing, there was a marginally significant negative main effect of inoculation (p=0.070), but no
significant interactions with emotionality or veracity.
For trustworthiness, the inoculation did not increase emotion discernment (p=0.583), and surprisingly
significantly reduced truth discernment (p = 0.036, d = 0.16).
Finally, for manipulativeness, there is a significant 3-way interaction (p=0.020) such that the inoculation
increases perceived manipulativeness of high emotion true and false content, as well as false low-emotion
content (d = 0.13).
 Experiment 2 - Sharing results
The inoculation video significantly increased sharing intentions (p=0.005) but had no other effects (e.g.
did not significantly change emotion discernment, p=0.316, or truth discernment, p=0.958).
 Experiment 3 with sample from MTurk

Methods

Methods were identical to Experiment 3 (as reported in the main text). Participants (N = 1,023 from
MTurk) were randomized to receive either the inoculation video or a control video. They rated the
perceived accuracy of the headlines, which were varied within-subject, with false headlines always being
high-emotion and true headlines always being low-emotion

Results
The inoculation had no significant main effect or interaction with veracity. However, the main effect of
veracity (although statistically significant, p=0.734) was much smaller in magnitude than observed in any
other analogous experiments (either in this paper, or the literature more generally). Thus, we concluded
that there was some idiosyncratic problem with this sample of respondents, and switched to Prolific for
the remaining experiments.