algorithms

Review

A Review of Explainable Artificial Intelligence from the

Perspectives of Challenges and Opportunities
Sami Kabir 1, *,†,‡ , Mohammad Shahadat Hossain 2,‡ and Karl Andersson 1, *,‡

1 Department of Computer Science, Electrical and Space Engineering, Luleå University of Technology,
SE-931 87 Skelleftea, Sweden
2 Department of Computer Science & Engineering, University of Chittagong, Chattogram 4331, Bangladesh;
hossain_ms@cu.ac.bd
* Correspondence: sami.kabir@ltu.se (S.K.); karl.andersson@ltu.se (K.A.);
Tel.: +46-705159568 (S.K.); +46-708195484 (K.A.)
† Current address: LGH 1309, Södra Lasarettsvägen 17, SE-93 132 Skelleftea, Sweden.
‡ These authors contributed equally to this work.

Abstract
The widespread adoption of Artificial Intelligence (AI) in critical domains, such as health-
care, finance, law, and autonomous systems, has brought unprecedented societal benefits.
Its black-box (sub-symbolic) nature allows AI to compute prediction without explaining
the rationale to the end user, resulting in lack of transparency between human and machine.
Concerns are growing over the opacity of such complex AI models, particularly deep learn-
ing architectures. To address this concern, explainability is of paramount importance, which
has triggered the emergence of Explainable Artificial Intelligence (XAI) as a vital research
area. XAI is aimed at enhancing transparency, trust, and accountability of AI models. This
survey presents a comprehensive overview of XAI from the dual perspectives of challenges
and opportunities. We analyze the foundational concepts, definitions, terminologies, and
taxonomy of XAI methods. We then review several application domains of XAI. Special
attention is given to various challenges of XAI, such as no universal definition, trade-off
between accuracy and interpretability, and lack of standardized evaluation metrics. We
conclude by outlining the future research directions of human-centric design, interactive
explanation, and standardized evaluation frameworks. This survey serves as a resource
for researchers, practitioners, and policymakers to navigate the evolving landscape of
Academic Editor: Ulrich Kerzel
interpretable and responsible AI.
Received: 13 June 2025
Revised: 27 August 2025 Keywords: accuracy; evaluation metrics; explainable artificial intelligence (XAI); human-
Accepted: 29 August 2025
centered design; interpretability; post hoc explanation; transparency; trust
Published: 3 September 2025

Citation: Kabir, S.; Hossain, M.S.;

Andersson, K. A Review of
Explainable Artificial Intelligence from
the Perspectives of Challenges and
1. Introduction
Opportunities. Algorithms 2025, 18, 556. The rapid advancement of Artificial Intelligence (AI) has led to transformative changes
https://doi.org/10.3390/a18090556 in a variety of fields, from healthcare and finance to autonomous systems and natural
Copyright: © 2025 by the authors. language processing. However, as AI systems become increasingly complex, the demand for
Licensee MDPI, Basel, Switzerland. transparency and interpretability in these systems has grown significantly [1]. In particular,
This article is an open access article the use of AI in safety-critical decision-making contexts has highlighted the importance
distributed under the terms and
of explainability [2]. AI models, particularly deep learning models, can achieve high
conditions of the Creative Commons
predictive accuracy in various tasks [3]. Despite their predictive prowess, these models
Attribution (CC BY) license
(https://creativecommons.org/
often function as “black boxes” due to the opacity of the internal reasoning process [4]. Such
licenses/by/4.0/). opacity has triggered the emergence of Explainable Artificial Intelligence (XAI) as a critical

Algorithms 2025, 18, 556 https://doi.org/10.3390/a18090556

Algorithms 2025, 18, 556 2 of 36

research area [5,6]. At its core, XAI refers to a set of techniques which provide human-
understandable explanation in support of the prediction of AI models [7]. This explanation
makes an AI model transparent and trustworthy. Such transparency is critical to ensure
regulatory compliance in sectors directly affecting individuals, such as finance, healthcare,
law, and criminal justice [8]. A major challenge of XAI is to make an AI model’s decision-
making process interpretable to humans, while maintaining high predictive accuracy [9].
Opaque deep learning models achieve state-of-the-art predictive accuracy in various
areas, such as image analysis, natural language processing, and disease prediction. How-
ever, such predictive results are not interpretable by humans [10]. This lack of transparency
is a significant barrier to the broader adoption of AI models in various safety-critical do-
mains. For instance, in the healthcare domain, disease prediction by an AI model has to be
both accurate and transparent [11]. Without transparency, the predictive output will not be
intelligible to doctors and patients, triggering a lack of trust in the AI decision [12]. Simi-
larly, in finance, regulators may require explanation for an AI-based credit scoring model’s
decision to check legal compliance and potential biases [13]. Moreover, AI explanation
can also teach humans new facts and knowledge. For example, Alpha Go Zero performs
better at a game of Go than its human counterpart [14]. If Alpha Go could explain its
gaming strategy, human players could learn new tactics to improve their proficiency. This
increasing need for transparency has driven the development of various XAI techniques to
improve the interpretability of AI models. For example, Local Interpretable Model-agnostic
Explanations (LIME) [4], a post hoc explanation technique, explains individual prediction
of a black-box AI model by approximating the prediction with a simpler, more interpretable
model. Another post hoc technique is SHapley Additive exPlanations (SHAP) [15], which
provides feature importance to explain a predictive output. Two primary categories of
XAI techniques are model-agnostic and model-specific approaches [16]. Model-agnostic
approaches can explain any machine learning model. Model-specific methods are tailored
to a specific machine learning model.
Despite making significant progress, the development of XAI techniques has to nav-
igate through several challenges. One major challenge is the trade-off in management
between model accuracy and interpretability [17]. Highly accurate deep learning mod-
els are difficult to explain due to their complexity and opacity [18]. In contrast, simpler
models, such as decision trees or linear regression, are easier to interpret. However, such
simple models may not provide as high predictive accuracy as a deep learning model
on complex tasks [19]. This trade-off between accuracy and interpretability is a central
dilemma in XAI, which causes users to accept reduced performance in exchange for more
interpretable models [20]. Moreover, a good explanation depends on the context in which
the AI model is deployed and on the audience receiving the explanation. For example,
a financial regulator’s need for explanation may be different than that of a healthcare
service provider [4]. Another significant challenge is the evaluation of the explanation [21].
Unlike conventional machine learning models’ metrics, such as accuracy and F1-score,
evaluating the quality of an explanation is inherently subjective and context-dependent.
Various metrics have been proposed, including fidelity (how well an explanation reflects
the actual decision-making process of the model) and sufficiency (whether the explanation
provides enough information for a user to make an informed decision) [22]. However, no
universally accepted framework exists for evaluating explanations of an AI model [23].
Moreover, the human factor plays a crucial role in the intelligibility of explanations [24].
Hence, explanation must be tailored to the audience’s level of expertise and cognitive
abilities. Otherwise, such explanation may cause misinterpretation of the model’s decision
by the audience [9]. For example, the explanation required by a data scientist is different
from that for a layperson [8]. Such subjectivity increases the complexity of XAI techniques.
Algorithms 2025, 18, 556 3 of 36

Additionally, biased or under-represented datasets pose the risk of misleading prediction

when using a machine learning model [25,26]. Hence, providing explanations to highlight
and address such biases is also a challenge for XAI techniques [27].
Despite these challenges, the field of XAI presents numerous opportunities for future
research and innovation. One promising direction is human-centered XAI, which focuses
on the human cognitive and psychological aspects to provide explanations. An explanation
can be made intelligible by utilizing insights from cognitive science and understanding
of human–computer interaction (HCI) [28]. For example, visual explanation, such as
heatmaps and saliency maps, can explain an image classification output by demonstrating
which parts of the image played the most influential role to provide this decision [29].
Another XAI approach is use of an interactive explanation interface, which allows users
to actively engage with the model’s decision-making process by asking questions or mod-
ifying inputs to evaluate the change in output [30]. Such interaction enhances a user’s
trust in an AI model, while providing valuable feedback to the developers to refine the
model further. Moreover, XAI can promote ethical AI by addressing fairness, accountability,
and transparency issues. Thus, by making an AI model interpretable, XAI can facilitate
the identification and mitigation of discrimination and biases present in AI decisions [31].
Finally, due to regulatory requirements, there is an increasing demand for AI systems to
comply with emerging laws and guidelines. For instance, the European Union’s General
Data Protection Regulation (GDPR) grants “right to explanation” to an individual for an
automated decision which affects the person significantly [32]. Such a legal framework cre-
ates an opportunity for XAI researchers to collaborate with policymakers and stakeholders
to ensure legal compliance of AI models.
The field of XAI is at a critical juncture as it seeks to balance the increasing power and
complexity of AI with the need for transparency, trust, and accountability. While significant
progress has been made in developing methods for explaining AI models, several challenges
remain, including trade-offs between interpretability and accuracy, lack of standardized
evaluation metrics, and understanding the human factors involved in providing meaningful
explanations. At the same time, exciting opportunities lie ahead in human-centered design,
interactive explanation interfaces, regulatory compliance, and ethical AI. This paper aims to
provide a comprehensive survey of these challenges and opportunities. Our goal is to bridge
the gap between technical advancements and practical needs, offering valuable insights
for researchers, practitioners, policymakers, and end-users. We begin by establishing
the foundational concepts and definitions in Section 2, followed by a taxonomy of XAI
techniques in Section 3. Section 4 highlights various application domains of XAI. Section 5
explores the practical and theoretical challenges impeding XAI implementation. Section 6
discusses the future scope of XAI. Finally, Section 7 concludes the paper.

2. Background and Definition of XAI

2.1. Background
The concept of explainability in AI has long been a topic of interest, dating back
to the development of early expert systems in the 1970s [33]. However, the modern
emergence of XAI as a formal research area is primarily driven by the increasing adoption
of complex opaque models, particularly deep learning models, and the growing demand
for transparency, accountability, and trust in automated decision-making systems.
Early Expert Systems and Symbolic AI: The roots of explainability in AI can be traced
back to rule-based expert systems, such as MYCIN, which was developed to diagnose
bacterial infections and recommend antibiotics [34]. MYCIN used a series of if-then rules
and provided human-interpretable explanations by tracing its inference chain [35]. Thus,
because of its symbolic reasoning structure, MYCIN had inherent explainability. However,
Algorithms 2025, 18, 556 4 of 36

symbolic systems have limitations in terms of scalability and uncertainty handling, leading
to a paradigm shift towards data-driven machine learning models [36].
The Rise of Black-Box Models: The paradigm shift from symbolic AI to statistical
and sub-symbolic machine learning resulted in the rise of more powerful predictive models,
such as random forests, Support Vector Machines (SVMs), and ultimately Deep Neural
Networks (DNNs) [37]. These models have achieved superior performance across various
application domains. However, the internal logic and feature importance of these models
are hidden from end-users, making them “black-boxes” [20]. The term “black-box” is
emblematic of the trade-off in these models between explainability and accuracy. The
increasing complexity and accuracy of these models come at the cost of reduced inter-
pretability. Such opaqueness is a critical issue in domains where transparency is essential
for safety, compliance, and user acceptance.
Deep Learning and the XAI Imperative: The widespread success of deep learning
in areas such as disease prediction, computer vision, and natural language processing has
intensified concerns about opacity and transparency [38]. DNNs consist of millions of
parameters and non-linear transformations, rendering their internal mechanisms nearly
incomprehensible to humans [18]. Deployment of such models in safety-critical areas, such
as healthcare, finance, and autonomous driving, exposes critical risks. To address these
risks, researchers seek to extract meaningful explanations from complex models. This type
of investigation has triggered the emergence of XAI as a new research area [8].
Institutional and Regulatory Factors: As an institutional initiative to promote ex-
plainability in AI, the Defense Advanced Research Projects Agency (DARPA) of the U.S.
Department of Defense launched an XAI program [5,6] in 2016 to create explainable ma-
chine learning models while maintaining a high level of prediction accuracy. This program
played a foundational role in formally defining XAI as a research domain and fostering
interdisciplinary collaboration. The Fairness, Accountability and Transparency (FAT) collab-
oration [39] is another group, which is focused on promoting explainability and reducing
bias in automated decisions produced by an AI model. Moreover, the European Union
(EU) has enacted a regulatory framework called the “General Data Protection Regula-
tion (GDPR)” which emphasizes the need for explainability of AI models.This regulation
introduced the “right to explanation” to individuals with respect to an automated deci-
sion which significantly concerns the person [32]. This legal mandate has enhanced the
importance of transparency of AI models.
Multi-disciplinary Research Field: In addition to algorithmic methods, the human-
centric dimension is taken into account by XAI. Hence, XAI is a multi-disciplinary research
field encompassing machine learning, HCI, philosophy, and cognitive psychology [40]. Re-
searchers argue that a good explanation has to be understandable, relevant, and actionable
by different users [23]. Consequently, XAI intersects with broader concerns of fairness,
accountability, usability, and societal impact. We show the timeline of the transition from
early expert systems to the present XAI in Table 1.

Table 1. Timeline of transition to XAI.

Year Milestone Description

One of the first AI systems to provide rule-based, human-readable
1975 MYCIN expert system [35]
explanations in medicine.
Growth of symbolic AI and Explainability was inherently part of early symbolic AI using logical
1980s
rule-based models [34] inference and transparent rules.
Transition to statistical machine Rise of machine learning models (e.g., random forests, SVMs),
1990s
learning [37] with reduced inherent interpretability.
2012 Deep Learning [41] Despite high predictive accuracy, DNN is an opaque black-box model.
Algorithms 2025, 18, 556 5 of 36

Table 1. Cont.

Year Milestone Description

U.S. government’s initiative to develop interpretable machine learning
2016 DARPA XAI Program [5]
models without compromising accuracy.
A post hoc explanation tool which uses local surrogate model to explain
2016 LIME [4]
individual prediction.
A post hoc tool, which uses Shapley values to identify
2017 SHAP [15]
feature importance.
“Right to Explanation” in The EU established legal grounds for demanding explanations of
2017
GDPR [32] automated decisions.
Highlights the importance of social science and human-centered
2019 Human-centered XAI [23]
evaluation in explanation design.
2020s Interdisciplinary research [40] XAI intersects with machine learning, HCI, philosophy, and ethics.

2.2. Definition
XAI refers to a class of methods, techniques, and models which have the aim of
making the decision-making processes of AI models transparent, interpretable, and un-
derstandable to human users. The increasing deployment of machine learning models,
particularly DNNs, has intensified the demand for transparency and interpretability [5,8].
XAI addresses the opacity or “black-box” nature of complex AI models. Although these
black-box models, such as DNNs and ensemble methods, offer high performance, lack
of intelligibility makes it difficult for end-users to understand how or why a particular
decision was made [20]. XAI addresses this limitation by either employing inherently
interpretable models or generating post hoc explanations which approximate the model’s
decision logic without changing the model’s structure [4,15].
U.S. DARPA, one of the early proponents of the term XAI, defines XAI as “an initiative
to produce more explainable models while maintaining a high level of learning performance
(prediction accuracy), and to enable human users to understand, appropriately trust,
and effectively manage the emerging generation of artificially intelligent partners” [5].
This definition emphasizes both the technical and human-centered goals of explainability,
while supporting user trust and effective interaction. Other researchers have expanded
this definition by distinguishing between interpretability, which refers to the degree to
which a human can consistently predict the model’s output, and explainability, which
encompasses the broader process of generating human-understandable reasons for the
model’s decisions [9,23]. Doshi-Velez and Kim [8] argue that XAI should not be narrowly
defined by the provision of explanation alone. They place emphasis on how well an
explanation serves the needs of different stakeholders, such as developers (for debugging),
end-users (for trust), and regulators (for compliance). Consequently, a universally accepted
definition of XAI remains elusive, as explainability is often context-dependent and tailored
to the goals, expertise, and concerns of the target audience. To make an explanation
meaningful, different domains have different requirements. We provide a comparative
overview of how XAI is interpreted across various domains in Table 2. In light of the
diversity of interpretations, we define XAI as follows:
XAI refers to a set of methods and frameworks which make the decisions of AI models
transparent, interpretable, and understandable to human stakeholders, with the goal of
enhancing trust, accountability, fairness, and regulatory compliance without significantly
compromising the AI models’ performance.
Based on this definition, we demonstrate how XAI methods make AI models transpar-
ent and trustworthy to human users in Figure 1. This definition forms the foundation for
examining the challenges and opportunities of XAI. Moreover, several terms in XAI have
ambiguous meanings, which we clarify in the next subsection.
Algorithms 2025, 18, 556 6 of 36

Table 2. Comparative interpretation of XAI across various domains.

Domain Interpretation Primary Goals

XAI is a framework, which produces more
Military and Government [5] explainable models while maintaining Trust, control, and operational reliability.
performance and enhancing user trust.
Explanation should provide clinically sound
Transparency, clinical trust, informed
Healthcare [12] reasoning, which physicians and patients
decision-making.
can understand.
Explanation should place emphasis on
Finance [13,32] regulatory compliance and bias detection in Fairness, accountability, legal conformance.
decision, such as credit scoring.
Focus is on the right to explanation for
Legal and Regulatory [32] algorithmic decisions which affect Interpretability for audits.
individuals.
Explanations assist in model debugging and
Developers/Engineers [4,15] Debugging, model improvement
feature importance analysis.
Explanations must be intuitive, simple,
End-users (Non-experts) [23,28] Understandability, user trust, adoption.
and support trust and usability.
XAI should systematize interpretability
Academic Researchers [8,20] through formal definition and Scientific rigor, reproducibility, generalization.
evaluation metrics.

Figure 1. AI model predicts output for the human user: (a) without explanation, resulting in lack of
trust; and (b) with explanation from an XAI method, resulting in trustworthy output.

2.3. Terminology
The field of XAI encompasses a diverse and often inconsistent set of terms used across
various disciplines. Hence, to ensure clarity and facilitate interdisciplinary collaboration, it
is important to define the core terminology of XAI. We provide an overview of the most
commonly used terms of XAI below.
Explanation: An explanation in XAI refers to the information provided by a model
which makes the model’s decision-making process comprehensible to humans. According
Algorithms 2025, 18, 556 7 of 36

to [23], explanations are social constructs, which should be aligned with the way humans
naturally seek and evaluate explanations. In the context of machine learning, an explanation
is defined as an interface between a human and AI to clarify how the inputs relate to the
outputs [8].
Interpretability: The degree to which a human can understand the internal reasoning
or parameters of a machine learning model is called interpretability [20]. Despite often being
used interchangeably with “explainability,” the term “interpretability” is more narrowly
focused on the transparency of the model’s internal structure. For example, decision trees
are considered interpretable because of the transparency of the decision paths [42].
Explainability: Explainability is a broader concept which encompasses both inter-
pretability and post hoc explanation [9]. To promote explainability of an opaque AI model,
various post hoc tools, such as SHAP [15] and LIME [4], are employed.
Transparency: The openness of a model’s structure and parameters is called trans-
parency [43]. A transparent model’s inner working logic is accessed, inspected, and un-
derstood by the relevant stakeholders [20]. Examples of transparent models include linear
regression and decision trees. DNN, by contrast, is non-transparent due to its opaque
internal structure [44].
Faithfulness: The extent to which an explanation properly reflects the actual reasoning
of a model is called faithfulness, also known as fidelity [45]. A faithful explanation does
not misrepresent the model’s actual decision-making process, even if the explanation is
simplified for human intelligibility.
Simulatability: A human’s ability to simulate or replicate a model’s decision-making
process is called simulatability [8]. A model’s simulatability is similar to interpretability,
with additional emphasis on the cognitive load of a human.
Justifiability: Ethical and social acceptance of the explanation of a model is called
justifiability [46]. Even if an explanation is technically correct, it may be unjustifiable due
to bias or discrimination. Such justification is important in certain critical areas, such as
criminal justice and healthcare.
Post Hoc Explanation: An explanation provided by various post hoc (after the
event) tools is called a post hoc explanation [47]. Such post hoc tools, applied after a
model’s predictive output, explain the model’s decision without revealing its internal
logic. Examples of post hoc explanations include feature importance scores, counterfactual
examples, and visualizations [4,15].
Inherently Interpretable Models: Such models are designed from the ground in a
transparent manner for human interpretability [48]. Examples of inherently interpretable
models include decision trees, linear models, and rule-based models [12]. The inherent
interpretability of such models often comes at the cost of accuracy. However, these models
still remain critical in domains where the transparency of decisions is non-negotiable.

3. Research Methodology
Several journals and databases provide a rich source of literature on XAI’s role to
make AI models’ decisions transparent. We conducted a systematic search through this
rich literature to identify reliable studies from credible authors published from 2022 to 2025.
As a verified framework to perform this search in a scientific, reproducible, and transparent
approach, we employed the Systematic Literature Review (SLR) [49] approach using the
PRISMA guidelines [50]. The SLR process consists of three phases: planning the review,
conducting the review, and reporting.
Algorithms 2025, 18, 556 8 of 36

3.1. Planning the Review

We undertook extensive reading and consulted with peers engaged in similar studies
to ensure full compliance with the steps required for SLR. This enabled us to minimize
the risk of unplanned research duplication and to maximize transparency of the literature
review. Special emphasis was placed on research ethics, proper citations, and the intellectual
property rights of other researchers. To guide the literature review, we formulated three
research questions, as shown in Table 3. With these questions, we systematically identified
the existing XAI methods in the literature, their application domains, and limitations.
To perform the literature search, we selected two databases, Scopus and Web of Science
(WOS), because of their extensive coverage of peer-reviewed XAI articles published in
leading international journals and conference proceedings [51]. To collect relevant articles
from these two databases, we formulated search queries by combining keywords related to
XAI. We show the search strings in Table 4.

Table 3. List of the research questions.

Research Question Description

What are the key XAI methods for making the decisions of AI
RQ1
models transparent?
How does the relevance of XAI methods vary based on the appli-
RQ2
cation domains?
What are the major limitations and challenges of XAI methods for
RQ3
enhancing the transparency of AI models?

Table 4. Search string.

Keywords Related to Explainability Keywords Related to AI

(“explainable” OR “interpretable” OR
“transparent” OR “trustworthy” OR “xai” (“artificial intelligence” OR “AI” OR
OR “fair” OR “ethical” OR “robust” OR “machine learning” OR “ML” OR “deep
“accountable” OR “bias free” OR learning” OR “neural networks” OR
“discrimination free” OR “safe” OR “automated decision making” OR
“post-hoc explanation” OR “post-hoc “predictive analytics” OR “algorithm”)
interpretability”) AND

3.2. Conducting the Review

Using the search queries, a total of 3173 papers were identified. To ensure align-
ment with our research objectives, each paper was assessed based on its title, abstract,
and keywords. We then applied inclusion and exclusion criteria as specified in Table 5
to filter relevant articles further for our review. We conducted the review in four primary
stages: identification, screening, eligibility determination, and sorting of research articles.
We represent each of these four stages in a PRISMA diagram [50], as shown in Figure 2.
In the identification stage, we searched the Scopus and WoS database for XAI articles from
2022 to 2025 using the keywords shown in Table 4. In the screening stage, we removed
duplicates and unrelated articles, resulting in 2425 unique articles. Out of these articles,
we determined that 122 articles were eligible based on the inclusion and exclusion criteria
of Table 5. In the sorting stage, we finally selected 101 articles for review. The PRISMA
diagram provides a transparent overview of the whole literature search process, which can
be used for replication of the survey.
Algorithms 2025, 18, 556 9 of 36

Table 5. Inclusion and exclusion criteria.

Inclusion Criteria Exclusion Criteria

XAI methods have been proposed to enhance
Technical reports and book chapters;
transparency of AI models;
Articles in English; Editorials, viewpoints, and opinions;
Novel contributions published in
peer-reviewed journals or conference Duplicated articles.
proceedings.

Figure 2. The PRISMA flowchart showing the stages of literature search, where n = number of articles.

3.3. Reporting
In this stage, we present the findings of the literature review. Of the 101 published
articles, 61 were published in peer-reviewed journals, and the remaining 40 in conference
proceedings. In terms of year-wise distribution, 41, 29, 22, and 9 articles were published
in the years 2022, 2023, 2024, and 2025, respectively. The number of articles published in
2024 and 2025 is lower due to indexing delays and use of partial-year data, respectively.
Among the keywords, ‘explainable artificial intelligence’, ‘machine learning’, and ‘deep
learning’ were the most prevalent. On the other hand, the least prevalent keywords were
‘transparent neural networks’, ‘trustworthy decision making’, and ‘post hoc interpretability’.
In terms of the geographical distribution, the highest number of publications were from the
United States (n = 26), followed by China (n = 21) and India (n = 11). Based on our findings
from the published articles included in the review, we present the taxonomy, application
domains, challenges, and future scope of XAI in the subsequent sections.

4. Taxonomy of XAI
A systematic taxonomy is essential to apply XAI methods across different domains
in an organized manner. Several taxonomic frameworks have been proposed, which
are generally categorized by explanation timing, explanation scope, model dependency,
explanation form, audience, and interaction modality [8,52,53]. We mention these XAI
classes below.
Algorithms 2025, 18, 556 10 of 36

Intrinsic (Ante Hoc) and Post Hoc Explainability: Based on explanation timing, XAI
techniques are classified into intrinsic (ante hoc) and post hoc methods [47].
• Intrinsic (ante hoc) methods are inherently interpretable models, such as decision
trees, rule-based models, and linear regression. These models offer transparency by
exposing the internal logic directly to the user [54]. For example, the Belief Rule-Based
Expert System (BRBES), a rule-based model with intrinsic explainabilty, was employed
by [25] to predict and explain the energy consumption of buildings.
• Post hoc methods, by contrast, apply interpretability techniques after a model’s pre-
dictive output. These techniques extract explanations from a complex “black-box”
model, such as DNNs and ensemble models. Examples of post hoc methods include
LIME [4], SHAP [15], Partial Dependence Plots (PDPs) [55], Individual Conditional
Expectation (ICE) Plots, counterfactual explanation [56], and anchors [57]. PDPs and
ICE plots are useful for understanding the relationship between a feature and the
predicted outcome, particularly for models that have complex feature interactions.
PDPs show how a predicted output changes with regard to the variation in a single
feature, keeping other features constant. ICE plots show the effect of a feature on
individual instances [58]. Anchors are if-then rules, which “anchor” a prediction. If the
conditions in a rule are true, the AI model will make the same prediction, even when
other features change. These post hoc tools provide valuable insights into an opaque
model’s decision-making process. For instance, Gradient Boosting Machine (GBM)
was used to predict the sepsis risk of Intensive Care Unit (ICU) patients [59]. The post
hoc tool SHAP explained this prediction by showing that the features, such as serum
lactate level, respiratory rate, and Sequential Organ Failure Assessment (SOFA) scores,
were the significant contributors to the model’s prediction [59]. Such explanations
provide clinicians with actionable insights into individual patient’s risks.
Global and Local Explanation: Based on the scope of explanation, XAI techniques
are classified into global and local explanation [60].
• Global explanation describes the overall behavior of a model across the entire input
space, allowing insights into feature importance and model structure [53]. For example,
SHAP can be used to produce global explanations by aggregating local (per-instance)
Shapley values across several predictions, typically using the mean (or mean abso-
lute) Shapley value to estimate each feature’s overall importance [61]. This global
explanation of SHAP was applied to a random forest model trained with electronic
health record data [62]. After predicting unscheduled hospital readmission with this
trained random forest model, the authors used SHAP values to produce a global
explanation by ranking features according to their overall influence across the entire
cohort. The global SHAP summary identified days of stay and age as the top two most
influential features. Such global explanation reveals risk factors for the whole system,
rather than individual patients [62].
• Local explanation focuses on the reasoning behind individual prediction. For example,
LIME explains the individual predictions of a complex model by creating a local
surrogate model (e.g., linear regression or decision tree), which mimics the behavior of
the complex model near the input space [4,15]. Anchors are applied locally around an
instance to capture the conditions of an if-then rule [57]. To provide local explanation
of individual test cases of breast masses classification output, LIME was applied to
a DNN by [63]. For a particular patient, LIME highlighted high values of texture,
smoothness, and concave points as determining factors for a malignant prediction [63].
Such case-level local explanation helped clinicians to inspect the specific reasoning
behind a single prediction, and increased trust in the model-assisted diagnosis.
Algorithms 2025, 18, 556 11 of 36

Model-specific and Model-agnostic Techniques: Based on model dependency, XAI

techniques can be classified into model-specific and model-agnostic approaches [64].
• Model-specific approaches are tailored to specific types of models. For instance, atten-
tion maps are suitable for deep learning in computer vision and Natural Language
Processing (NLP) tasks [65], whereas feature importance scores are applicable to tree-
based models. TreeSHAP, a model-specific XAI technique for tree-ensemble models,
leverages the internal structure of decision-tree ensembles to compute exact Shapley
feature attributions efficiently. In a real-world study, TreeSHAP was employed by [66]
to explain a random forest model trained on a clinical metabolomics dataset. Tree-
SHAP produced both local and global feature importance explanations and identified
testosterone metabolites as a key discriminator in a urine dataset [66].
• Model-agnostic approaches are applied universally to any AI model to extract expla-
nations. Examples of these approaches include LIME, SHAP, and partial dependence
plots [67]. KernelSHAP is a model-agnostic estimator of Shapley values [68]. It treats
a trained predictor as a black-box model and approximates each feature’s contribution
by querying the model on perturbed inputs. In a clinical application to intelligent
fetal monitoring, ref. [68] applied KernelSHAP post hoc to a stacked-ensemble clas-
sifier to locally explain each instance of cardiotocography prediction. To facilitate
clinical interpretation, the authors also aggregated local importance scores to produce
global explanations.
Type of Explanation Output: The form of explanation is critical to its effectiveness.
Based on format, XAI can be classified into the following three types [69].
• Feature-based Explanation: This provides explanations by highlighting the contribu-
tion of each input feature to the prediction [70]. In a real-world study, ref. [71] applied
a local feature-based explainer (LIME in Study 1, and SHAP in Study 2) to predict
real-estate price. The authors showed that the feature-level explanations systematically
changed users’ situational information processing and mental models.
• Example-based Explanation: This uses training instances to explain a prediction
by highlighting its decision boundaries. Moreover, counterfactuals are generated to
inform the user of preconditions to obtain an alternative outcome [56]. Thus, it offers
intuitive “what-if” information at the instance level. In a real-world study, ref. [72]
introduced GANterfactual, an adversarial image-to-image translation method, which
generates realistic counterfactual chest X-rays to flip a pneumonia classifier’s decision.
The authors showed that these counterfactuals improved non-expert users’ mental
models, satisfaction, and trust compared with saliency map explanation.
• Visual Explanation: This technique (saliency/pixel-attribution maps) produces
heat-maps over images to highlight the regions the model used to form a prediction.
For example, Gradient-weighted Class Activation Mapping (Grad-CAM), a saliency
map technique, produces a heatmap to show which regions of the image contributed
the most to the predicted class [73]. If a Convolutional Neural Network (CNN) predicts
“dog” for an image, Grad-CAM may highlight the dog’s face and tail, indicating that
these regions were the most influential to the classification. In a real-world study,
ref. [74] systematically evaluated seven saliency methods (Grad-CAM, Grad-CAM++,
Integrated Gradients, Eigen-CAM, Deep Learning Important FeaTures (DeepLIFT),
Layer-wise Relevance Propagation (LRP), and occlusion) on chest X-rays. This study
found better performance of Grad-CAM than other saliency methods to localize
pathologies. However, all methods still performed substantially worse than a human
radiologist benchmark. This demonstrates that saliency heat maps are useful but not
fully reliable as a standalone clinical explanation [74].
Algorithms 2025, 18, 556 12 of 36

Expert and Layperson Orientation: Based on the intended audience, XAI techniques
are divided into expert-oriented and layperson-oriented explanation [75].
• Expert-oriented explanation provides detailed technical insights tailored to the devel-
opers, data scientists, or domain experts [25,26]. For example, ref. [76] developed a
modified Progressive Concept Bottleneck Model, which returns both anatomical seg-
mentations and clinically meaningful property concepts (symmetry, caliper-placement
feasibility, image quality) as real-time feedback for fetal growth scans. The system was
validated across hospitals, where 75% of clinicians rated the explanation as useful. More-
over, the model achieved 96.3% classification accuracy on standard-plane assessment.
• Layperson-oriented explanation is provided in non-technical language using analo-
gies, narratives, or visuals to enhance public understanding and trust [8]. Recent
advancements in XAI have increasingly focused on designing explanations which
align with human cognitive processes. To present explanations to a layperson in an
intelligible manner, principles for explanatory debugging were proposed by [28]. In a
real-world study, ref. [71] tested feature-level, plain explanations with lay participants
on a real-estate price prediction task. The study demonstrated that such explana-
tions change lay users’ decision patterns and mental models, highlighting both the
usefulness and risks of exposing non-experts to model rationales. Interactive natural-
language explanatory interfaces, such as TalkToModel [77] also increase non-expert
users’ performance and comprehension in applied settings.
Static and Interactive Explanation: Based on interaction modality, XAI techniques
are classified into static and interactive explanations [78].
• Static explanation provides fixed reports or visualizations without user input [79].
For example, PDPs are static, one-shot visual explanations to show the average ef-
fect of a feature on model predictions. PDPs are produced once and inspected as
non-interactive figures. In a real-world study, ref. [80] applied one-way and multi-
way PDPs to a gradient-boosted model for satellite-based PM2.5 prediction. PDPs
visualized how meteorological and spatiotemporal predictors influenced predicted
pollution levels across regions and seasons. Thus, the authors communicated the
model behavior to domain scientists using the static PDP figures.
• Interactive explanation allows a user to interact with the model. Such explanation
enables a user to investigate model behavior dynamically by changing inputs and ex-
ploring “what-if” scenarios, resulting in deeper comprehension of the model’s decision-
making process [81]. For example, an AI-powered recommender system can allow
users to adjust preferences and immediately assess how these changes influence
the recommendations [30]. This interactivity increases user trust and improves the
overall understanding of AI behavior. A prominent example of interactive explana-
tion in practice is the InteraCtive expLainable plAtform for gRaph neUral networkS
(CLARUS) [82]. This is an explainability platform for graph neural networks in clinical
decision support systems. CLARUS visualizes patient-specific biological networks
and model relevance scores. It also allows domain experts to manually edit graphs
(nodes/edges) to ask “what-if” counterfactuals, immediately re-predict outcomes,
and retrain models to observe the consequences of those edits. This interactive loop
moves beyond static saliency or post hoc attributions by allowing users to probe
causal hypotheses and refine model behavior. This interactive pattern is increasingly
emphasized in the human-centered XAI literature [83].
The taxonomy provides a structured overview of various classes of XAI techniques,
as shown in Figure 3. A comparison among different XAI techniques is presented in Table 6.
Various software frameworks and libraries related to XAI techniques, along with their
Algorithms 2025, 18, 556 13 of 36

model compatibility, applicable AI models, key features, and application domains are
presented in Table 7. To ensure maximum transparency, choice of an XAI technique must
be aligned with the requirements of the stakeholders concerned and the application context.
We focus on some of the application domains of XAI in the next section.

Figure 3. Taxonomy of XAI methods.

Table 6. Comparison among various XAI techniques.

Explanation Model
Technique Type Advantages Limitations
Format Compatability
Feature Interprets any Unstable
LIME Post hoc, Local Importance Model-agnostic model; good local explanations; lacks
(weights) fidelity. global consistency.
Solid theoretical Computationally
Post hoc,
SHAP Shapley Values Model-agnostic foundation; intensive for
Local and Global
consistent. complex models.
Intuitive for image Sensitive to noise;
Deep Learning
Saliency Maps Post hoc, Local Visual (Heatmap) data; visual lacks standardiza-
(CNNs)
feedback. tion.
Simple
Graphical Assumes fea-
PDP Post hoc, Global Model-agnostic visualization of
(Feature vs. Output) ture independence.
feature impact.
Reveals Difficult to
Graphical
ICE Post hoc, Local Model-agnostic heterogeneity in interpret in high
(Instance-level)
predictions. dimensions.
Offers actionable, Hard to generate
Counterfactuals Post hoc, Local Example-based Model-agnostic intuitive for complex
explanations. models.
High precision; Narrow coverage;
Rule-based
Anchors Post hoc, Local Model-agnostic human-readable computationally
(If-Then)
rules. expensive.
Global
understanding; Oversimplifies
Surrogate Models Post hoc, Global Tree/Rule-Based Model-agnostic
interpretable complex models.
model.
Built-in
Attention interpretability; May not faithfully
Intrinsic, Local Weighted Inputs DNN
Mechanisms aligns with human reflect reasoning.
focus.
Not
Feature Quick insights on instance-specific;
Post hoc, Global Ranked Features Model-agnostic
Importance key features. may miss fea-
ture interactions.
Algorithms 2025, 18, 556 14 of 36

Table 7. Software frameworks and libraries of XAI techniques.

Software
Model Applicable AI Application
Framework and Technique Key Features
Compatibility Models Domains
Library
Human-readable
Local surrogate
local explanation, Disease prediction,
models, Feature- Tabular classifiers
LIME [84] Model-agnostic fast, lightweight, financial
importance and regressors.
pluggable to any fraud detection.
weights.
predictor.
Contrastive
Shapley-value Fairness auditing
analysis, supports
attributions, Tree ensembles, in safety-critical
SHAP [85] Model-agnostic batching and
Global and local DNN. domains, biomedi-
kernel
explanations. cal diagnosis.
approximations.
Integrated
gradients, Saliency
maps, Deep Extensible API, Attribution for
Learning multi-modal Large Language
PyTorch models CNN, RNN,
Captum [86,87] Important support, tight Models (LLMs),
(version 2.1.0) Transformer.
FeaTures PyTorch NLP
(DeepLIFT), integration. model debugging.
SmoothGrad,
VarGrad.
Simple API, Clinical decision
Permutation
built-in support systems,
importance, Decision trees,
ELI5 (Python Scikit-learn visualization, production
Decision tree Linear models,
library) [88] estimators produces machine learning
weight extraction, Random forests.
human-readable pipeline debug-
Text explanations.
explanations. ging.
Counterfactual Industrial Internet
Decision trees, Built-in evaluation
explainers, of Things (IoT)
Model-agnostic Random Forests, metrics, taxonomy
AI Explainability Contrastive forecasting,
and model-specific Logistic guidance,
360 (AIX360) [89] methods, TS-LIME, anomaly detection,
modules Regression, SVM, plug-and-play
TS-SHAP for supply
DNN, RNN. pipeline.
time-series. chain analytics.

5. Application Domains of XAI

XAI has found applications across various domains, each with unique challenges and
requirements for transparency and interpretability. Understanding the different contexts in
which XAI is applied is crucial for identifying the specific demands of each domain and for
tailoring XAI methods to ensure effective use. This section outlines several key application
domains of XAI, highlighting the importance of explainability in these fields.

5.1. Healthcare
In healthcare, AI models have been increasingly utilized for various tasks, such as
medical diagnosis, personalized treatment plans, and drug discovery. These models,
particularly DNNs, can achieve high predictive accuracy. Disease prediction by DNNs
concerns life and death questions for patients. Lack of interpretability of disease prediction
raises concerns for patients’ safety [12]. Hence, the explainability of an AI model’s decision
in healthcare is of paramount importance. This explainability enables medical professionals
to evaluate whether the decision of an AI model is consistent with medical knowledge and
ethics [90]. In this context, XAI methods are used to explain the decision of medical AI,
resulting in a trustworthy predictive output [91].
Algorithms 2025, 18, 556 15 of 36

For instance, cancer diagnosis from Magnetic Resonance Imaging (MRI) images using
an AI model has to be both accurate and transparent [12]. For transparency, an XAI method
saliency map can be employed to highlight the important features of an MRI image which
contributed the most to detect cancer [15]. We focus on some key aspects of the healthcare
domain from the XAI perspective below.
• Key AI models and XAI methods
– Diagnostic Imaging: CNN with Grad-CAM [92].
– Disease Prediction: Ensemble tree algorithm with SHAP value explanation [93].
– Treatment Recommendation: Rule-based model [94].
• Domain features
– Highly sensitive, and heterogeneous data, such as imaging, time-series data of
Electronic Health Record (EHR), and genomics [92].
– Lack of adequate quantity of labeled data. [93]
– Erroneous prediction and explanation have severe consequences. Hence, the
healthcare domain needs transparent prediction from AI models with high accu-
racy [94].
• Problem types
– Classification: Binary (disease versus no disease) and multi-class classifica-
tion [95].
– Regression: Risk score prediction for a patient, such as readmission probability,
and mortality risk [96].
– Sequence-to-sequence: Clinical report summarization [94].
• Advantages of using XAI
– XAI improves clinicians’ trust in AI decisions through feature importance scores
or example-based explanations, such as SHAP explanation in support of a heart
failure prediction model [96].
– XAI facilitates regulatory compliance, such as “right to explanation” of GDPR [94].
• Disadvantages of using XAI
– Explanations offered by various post hoc XAI techniques may be misleading due
to local approximation of a decision boundary instead of capturing the full model
logic or global behavior [97].

5.2. Finance
AI models are used extensively in finance for tasks such as credit scoring, fraud
detection, algorithmic trading, and risk management [98]. However, the “black-box” nature
of many AI models, especially DNNs, presents challenges in ensuring transparency and
accountability of the AI decisions [13]. Hence, the explainability of such decisions is critical
in the financial sector to mitigate financial and personal consequences. Regulators and
institutions require clear explanations for AI-driven decisions to ensure fairness, detect
biases, and to comply with financial regulations [99].
For instance, a model-agnostic approach SHAP [15] can be employed to explain the
decision of a credit scoring AI model by identifying the features which contributed the
most to a low credit score [100]. We focus on some key aspects of the finance domain from
an XAI perspective below.
• Key AI models and XAI methods
– Credit scoring: Random forest with SHAP value explanation [101].
– Financial time-series forecasting: Recurrent Neural Network (RNN) with SHAP
value explanation [102].
Algorithms 2025, 18, 556 16 of 36

• Domain features
– High compliance requirements: Financial AI models must be fully auditable
and traceable to meet standards such as GDPR, Basel-III, and the Fair Credit
Reporting Act [101].
– Fairness and Non-Discrimination: Bias in financial data, such as in relation to
gender, race, and location, poses legal and ethical risks [103].
– Real-time requirements: Applications such as fraud detection and high-frequency
trading need sub-second predictions and lightweight explanation methods (e.g.,
SHAP, surrogate decision trees) to keep pace with streaming data [103].
• Problem types
– Classification: Financial fraud and anomaly detection as normal versus suspi-
cious [104].
– Regression: Stock price forecasting and credit scoring by predicting the proba-
bility of default [105,106].
– Optimization: Investment portfolio optimization through asset allocation [107].
• Advantages of using XAI
– Feature importance and counterfactual explanations enable financial analysts to
address demographic or gender biases in credit scoring and lending models [16].
– Explanations provided by XAI methods enable financial analysts to improve their
decision-making speed and quality [108].
• Disadvantages of using XAI
– When faced with large high-frequency datasets of trading and fraud-detection
pipelines, traditional post hoc XAI techniques, such as LIME and SHAP, may
produce overly generalized or even misleading explanations [109].
– Explanations may leak the sensitive proprietary logic of a financial AI model,
which can be manipulated to fool the AI model [109].

5.3. Criminal Justice

AI models are increasingly being used in criminal justice systems for predictive
policing, sentencing decisions, and parole evaluation [110]. However, the use of opaque
AI models in such critical decision-making areas raises serious concerns about fairness,
bias, and accountability [100]. In this context, XAI methods are essential to ensure that the
AI decisions are bias-free and justifiable to the stakeholders, including judges, lawyers,
and the public.
For instance, in predictive policing, a police department may use a DNN to predict
the likelihood of crime in different neighborhoods based on relevant features, such as
historical crime data, time of day, location demographics, and economic indicators [110].
However, as a DNN is an opaque model, officers and policy makers will not understand
why a specific area is flagged as high-risk by the DNN. In this case, the post hoc tool
LIME can be employed to demonstrate which features of the locality most influence a
certain prediction regarding crime risk [111]. Similarly, to explain an AI-driven legal
sentence, SHAP can be employed to highlight the most influential features contributing to
the sentence. Such transparency reduces discrimination and supports more equitable legal
practices [27]. Several key aspects of the criminal justice domain from the XAI perspective
are indicated below.
Algorithms 2025, 18, 556 17 of 36

• Key AI models and XAI methods

– Recidivism prediction: Gradient-boosted decision trees (XGBoost) with SHAP
dependence plots to reveal how factors such as age at first offense, incarceration
length, and in-prison behavior drive risk scores to re-offend [112].
– Spatial crime forecasting: Random forest predicts burglary or street-crime risk
patterns, which are explained by SHAP values [113].
• Domain features
– Fairness-critical decisions: Decisions such as bail release, sentencing, and pa-
role directly affect individual liberties. Hence, judicial AI models must be both
accurate and transparent [114].
– Severe class imbalance: Criminal events such as violent re-offense and homicide
cases are rare. Hence, skewed class distributions have to be addressed properly
by judicial AI models [115].
– Multi-modal inputs: Criminal data can be in the format of tabular records
(demographics, prior convictions), text (legal documents), biometrics (face, finger-
print), and imagery. Hence, judicial AI models have to deal with heterogeneous
multi-modal input data [116].
• Problem types
– Classification: Whether a criminal will recidivate or not has to be classified [117].
– Regression: Predicting the continuous number of crimes over time in a
hotspot [118].
– Sequence-to-sequence: Converting unstructured legal texts into structured
summaries [119].
• Advantages of using XAI
– Post hoc XAI techniques such as SHAP and LIME reveal which features drive risk
scores or case clearance predictions, enabling practitioners to detect and correct
biases embedded in historical data or AI models [120].
– Providing judges, attorneys, and defendants with comprehensible explanations for
AI-driven decisions fosters public confidence in the criminal-justice process [120].
• Disadvantages of using XAI
– Through explanations provided by XAI techniques, exploitable triggers of a judi-
cial AI model may be revealed. Such triggers may be manipulated by malicious
actors to reverse engineer the AI model and evade detection [121].
– Many post hoc XAI techniques produce simplified explanation,s such as
single-feature attribution, which may omit critical model complexities. Such
simplification may causes judges to interpret the AI model incorrectly [122].

5.4. Autonomous Systems

Autonomous systems such as self-driving cars and drones rely heavily on AI models
to make real-time decisions in complex, dynamic environments [123]. Decisions of au-
tonomous systems directly affect human safety. Hence, the transparency and interpretability
of such decisions is critical to ensure accountability [124]. To promote transparency, XAI
plays a crucial role by explaining these decisions made by autonomous systems, such as
why a car chooses to brake or swerve in a given situation. Thus, XAI can improve trust in
autonomous systems, ensure compliance with safety regulations, and aid in debugging the
systems for further improvement [125].
For instance, saliency map-based explanation can provide insights into the decision-
making process of an autonomous car to detect obstacles and navigate [46]. Such explana-
Algorithms 2025, 18, 556 18 of 36

tions can enable developers to verify whether the autonomous car acts rationally, especially
in critical situations. Several key aspects of autonomous systems from the XAI perspective
are indicated below.
• Key AI models and XAI methods
– Image segmentation: CNN performs pixel-wise semantic segmentation of road
scenes. Critical image regions such as pedestrians and lane markings can be
highlighted by saliency maps. Such highlighted regions facilitate design-time
debugging and runtime safety monitoring of self-driving cars [126].
– Anomaly detection: Random forest is applied on multi-sensor streams such as
cameras, Inertial Measurement Units (IMUs), and the Global Positioning System
(GPS) to detect anomalies or faults in a drone. Post hoc XAI techniques such as
SHAP and LIME can be applied to fuse and rank the most salient sensor-level
features [127].
• Domain features
– Real-time closed-loop operation: Autonomous agents deal with continuous
streams of heterogeneous sensor data. AI models have to act on these sensor data
within milliseconds. This “sense–think–act” loop requires an AI model to have
ultra-low latency for safety-critical actions [128]. Moreover, XAI techniques such
as saliency maps and confidence scores have to be generated without breaking
real-time constraints [128].
– Safety and reliability: In self-driving cars or industrial robots, failures can lead
to injury or loss of life. Hence, such systems must conform to functional safety
standards, such as ISO 26262, and provide stable explanations. Such explanations
enable engineers and regulators to inspect an autonomous system before and
after an incident [129]. Otherwise, without robust explanations of XAI methods,
loopholes in autonomous systems will remain undetected.
– Dynamic environments: Autonomous systems operate in non-stationary set-
tings, such as changing weather and where there are novel obstacles. Hence,
to cover this distribution shift, XAI methods have to convey uncertainty and
adaptation over time [130].
• Problem types
– Classification: Objects or events in the environment are classified by AI mod-
els [131]. For example, an AI model can classify obstacles as pedestrians, cyclists,
or vehicles. Similarly, a scene can be classified as a road sign or lane marking.
– Regression: AI models can regress continuous quantities for an autonomous
system, such as future positions, velocities, and risk scores [132].
• Advantages of using XAI
– Natural language explanations for driving commands produced by transformer-based
architectures enhance the trustworthiness of AI models [129].
– Post hoc XAI tools such as attention-map visualizations, surrogate models,
and feature-importance scores highlight which inputs triggered a particular
decision. This fine-grained insight enables developers to identify the architectural
weakness of an autonomous system [129].
• Disadvantages of using XAI
– Many post hoc XAI methods, such as LIME, SHAP, and saliency maps, entail high
computational costs to process high-dimensional sensor inputs. Such extensive
computation can cause delays in perception and planning pipelines of a real-time
autonomous system [128].
Algorithms 2025, 18, 556 19 of 36

– XAI explanations may become unstable or misleading when the sensor quality
degrades due to various reasons, such as low light, fog, rain, and motion blur.
Without properly quantifying such uncertainties, XAI techniques may provide
incorrect explanations in critical situations [133].

5.5. Customer Service and Human Resources

In customer service, AI-powered chatbots and recommendation systems are frequently
used to interact with customers and make personalized suggestions [134]. The explainabil-
ity of these systems is crucial to ensure that customers trust the recommendations they
receive. Similarly, in human resources, AI is increasingly being used for resume screening,
employee performance evaluation, and promotion decisions [9]. As these decisions directly
impact individuals’ careers, ensuring that the AI systems are transparent and free from bias
is critical [46,135]. We highlight some key aspects of customer service and human resources
from the XAI perspective below.
• Key AI models and XAI methods
– Chatbots and virtual assistants: In customer service, various Large Language
Models (LLMs), such as Generative Pre-trained Transformer (GPT) and Bidirec-
tional Encoder Representations from Transformers (BERT), are used to develop
chatbots and virtual assistants [136]. As an XAI method, attention visualization
is used to explain chatbot responses [136].
– Resume screening: As an AI model, random forest is applied by human re-
sources departments of organizations to screen resumes of candidates [137].
To explain the candidate selection, SHAP is applied [137].
• Domain features
– Real-time high-volume engagement: To serve customers efficiently, chatbots
must handle large streams of customer requests (e.g., chat, voice, and email).
Hence, XAI methods have to provide on-the-fly explanations to sustain user
engagement. [138].
– Multi-function coverage: In human resources, AI models are employed for
multiple functions, such as talent acquisition, performance appraisal, and workforce
planning. Hence, based on the use case, the algorithmic requirements of an AI model
can be of various types, such as descriptive, predictive, and prescriptive [139].
• Problem types
– Classification: Customer service departments can employ AI models to classify
whether a customer will churn [140]. Human resources can classify whether an
employee will stay or leave using AI models [141].
– Regression: A customer’s lifetime value can be predicted using a meta-learning-
based stacked regression approach [142].
– Clustering: By using AI models, human resources can cluster employees into
groups based on multi-modal features, such as skills, engagement, and perfor-
mance. Such clustering enables management to identify cohorts for tailored
training programs [143].
• Advantages of using XAI
– Transparent explanations of model outputs, such as why a recommendation
was made, significantly boost end-user confidence with AI-driven customer
services [136].
– Explainable candidate-screening systems reduce discriminatory outcomes by
making decision pathways auditable and transparent [137].
Algorithms 2025, 18, 556 20 of 36

• Disadvantages of using XAI

– Generating post hoc explanations for high-volume, real-time interactions can
introduce computational bottlenecks, slowing response time and degrading cus-
tomer experience [128].
– In human resources, the same explanation may be interpreted differently by other
managers, resulting in divergent decisions [71,144].

5.6. Large Language Models

With the advent of LLMs, such as GPT, the Pathways Language Model (PaLM), and the
Large Language Model AI (LLaMA), the landscape of AI applications has expanded dra-
matically [138]. LLMs enhance human–AI interaction through natural language generation.
Therefore, LLMs are being increasingly deployed in diverse domains, such as education,
healthcare, customer service, legal document analysis, and software development. How-
ever, the opacity of LLMs makes it challenging to understand how specific outputs are
generated [145]. To address this challenge, integration of XAI techniques with LLMs has
become a crucial research focus. For this purpose, various XAI techniques, such as attention
visualization [146], prompt attribution [147], explanation with exemplar-based reasoning
(e.g., in-context learning) [147], chain-of-thought prompting [148], and self-explaining
models [149] are being investigated to provide a transparent view of the model’s internal
logic [150,151]. Integration of these XAI techniques can contribute to a safer and more
trustworthy deployment of LLMs in high-stakes environments. However, more research is
needed to evaluate the reliability of such XAI techniques for transparency in LLMs [152].
Several key aspects of LLMs from the XAI perspective are highlighted below.
• Key AI models and XAI methods
– BERT is used as an LLM. As an XAI method, attention visualization highlights
which input tokens BERT attends to while making a prediction [147]. For example,
the attention visualization technique highlights keywords driving sentiment
classification.
– Another LLM is GPT-3. To explain the outcome of GPT-3, the XAI technique
Chain-of-Thought (CoT) is used [153]. CoT makes GPT-3’s logic transparent by
embedding intermediate reasoning steps directly with prompts.
• Domain features
– Model scale and architecture: LLMs are built on a huge-scale parame-
terized transformer architecture with self-attention, positional embeddings,
and multi-head attention layers [154]. This architecture enables long-range con-
text modeling of LLMs more effectively than using RNNs or Long Short-Term
Memory (LSTM) [154].
– Adaptation and human feedback: LLMs can be instruction-tuned to follow
diverse task descriptions. Moreover, LLMs can be further enhanced through
Reinforcement Learning from Human Feedback (RLHF) to refine responses and
align with human preferences [147].
• Problem types
– Classification: LLMs are fine-tuned for various classification tasks, such as
token-level classification, topic classification, and sentence classification [155].
– Generation: Decoder-only LLMs are trained for next-token prediction and used
for open-ended text generation, story generation, code generation, and other
creative tasks [155].
Algorithms 2025, 18, 556 21 of 36

– Structured output: LLMs support tasks requiring structured outputs, such as

relation extraction outputs in JSON format, particularly in medical and legal
information extraction settings [155].
• Advantages of using XAI
– By providing human-understandable explanations, XAI techniques enable stake-
holders to audit and sanity-check outputs of LLMs [156].
– LLMs can be prompted to generate their own explanations by approximating
traditional feature-attribution methods. Such explanations can be looped back to
fine-tune the factuality of LLM output [157].
• Disadvantages of using XAI
– Post hoc XAI techniques may miss complex interactions of transformer layers,
resulting in incomplete rationales. Hence, the end-user cannot understand the
true decision process of LLMs [128].
– Applying XAI techniques to LLMs is computationally expensive and latency-
intensive, which represents a serious bottleneck to provide real-time explanations
of LLM output [128].
Different AI models are used in different application domains. Some AI models are
used across multiple domains. In each of these domains, explainability is a precondition
to ensure trustworthy and justifiable use of AI models. In Table 8, we highlight major
AI models, their cross-domain applications, relevant XAI techniques, and each model’s
strengths and weaknesses. The increasing reliance of these domains on AI highlights the
importance of progressing XAI techniques to meet the specific requirements of each domain.
In the path to this advancement of XAI techniques, various trade-offs and challenges exist,
which we explain in the next section.

Table 8. AI models and XAI techniques across various domains.

AI Model Domain XAI Technique Strength Weakness

Decision Tree, credit scoring, inherently overfitting,
SHAP, surrogate tree
Random Forest consumer purchase interpretable, faster ensembles lose
visualization.
[158,159] modeling. to train. transparency.
global surrogate robust in high
clinical risk does not scale to
models, dimensions, clear
SVM [160,161] prediction, text large data, kernel
prototype-based margin
classification. sensitivity.
kernels. interpretation.
predictive probabilistic graph
uncertainty structured learning is
Bayesian maintenance, explanation,
quantification, hard to scale, domain
Networks [162,163] ecological counterfactual
causal reasoning. expertise required.
forecasting. inference.
feature visualization, high representation high computation
computer vision,
DNN [164] activation power, end-to-end cost, explanation
genomics.
maximization. learning. may lack fidelity.
social
edge importance leverages relational scalability issues,
Graph Neural recommendation,
scoring, activation structure, explanation
Network [165] traffic flow
masking. parameter-efficient. approximation.
prediction.
NLP, code attention-based modeling long-range opaque internals,
Transformer [146] generation, analysis, CoT dependencies, high
multi-modal fusion. rationales. in-context control. computation cost.
Algorithms 2025, 18, 556 22 of 36

6. Trade-Offs and Challenges of XAI

Despite the significant progress in XAI research, numerous trade-offs and challenges
persist, which complicate the practical deployment of interpretable AI models. In this
section, we explore these challenges and highlight the direction for XAI to evolve.

6.1. Accuracy Versus Interpretability Trade-Off

One of the most prominent challenges in XAI is the trade-off between model per-
formance and interpretability, often referred to as the accuracy–interpretability trade-off.
Complex models, such as DNNs, provide high predictive accuracy on a wide range of
tasks. However, due to their intricate and high-dimensional structure, interpretation of
this prediction is difficult. On the other hand, simpler models, such as decision trees and
linear regression, are easier to interpret. However, in many cases, the predictive accuracy
of these simple models may be lower than for complex models [12]. To bridge this gap be-
tween accuracy and interpretability, LIME [4] approximates a complex model with a locally
interpretable simple model. Similarly, SHAP [15] provides the global feature importance
of a black-box model. However, these XAI techniques cannot fully explain the internal
decision-making process of a complex model. The key challenge is to determine the right
balance between accuracy and interpretability for each application domain. For instance,
in high-risk domains, such as healthcare and finance, the interpretability of an AI model is
more critical than accuracy.

6.2. Evaluation Metrics

Unlike traditional machine learning metrics, such as accuracy, precision, or recall,
which are objective and well-defined, the evaluation of explanations involves inherently
subjective and context-dependent criteria [166]. Therefore, evaluation of XAI methods
remains a complex and multifaceted challenge. This evaluation concerns the faithfulness,
comprehensibility, usefulness, trustworthiness, utility, and interpretability of an explana-
tion. Several metrics have been proposed to assess the effectiveness of XAI methods, each
focusing on different aspects of explanation quality. One common metric is fidelity [167],
which measures how well an explanation approximates the decision-making process of the
original model. High fidelity ensures that the explanation reflects the true reasoning of the
model, making it a critical metric in the evaluation of local explanation methods, such as
LIME, ref. [4] and SHAP [15]. Stability is another key metric to measure how consistent
explanations are across different inputs or perturbations [168]. Human interpretability
of an explanation to end-users can be assessed using qualitative survey [8]. The level of
interpretability is determined by various factors, such as the explanation length, sparsity,
and linguistic simplicity. However, these factors are often heuristic and may not generalize
across different user groups [8]. To quantify the mental effort applied by an end-user to un-
derstand an explanation, a cognitive load metric is used [23]. Lower cognitive load results
in more effective interpretability of an explanation. Trust is also an important evaluation
metric in safety-critical domains, such as healthcare and finance [169]. To measure an end
user’s trust in a model’s decision after receiving explanations, quantitative surveys can
be conducted [8]. Finally, actionability metrics evaluate whether an explanation provides
sufficient information to end-users to make informed decisions, which is critical in sensitive
areas, such as medical diagnosis and judicial decisions [56].
The challenge in evaluating XAI lies in the development of a comprehensive frame-
work to capture the multi-dimensional nature of an explanation. Future research should
place emphasis on the standardization of these evaluation metrics to meet the diverse
requirements of various application domains [21]. Moreover, future XAI metrics can
evaluate how well explanations capture the true causal mechanisms, not just statistical
Algorithms 2025, 18, 556 23 of 36

associations [170,171]. The creation of large-scale, domain-diverse datasets with human-

annotated ground truth explanations can significantly improve the benchmarking of XAI
metrics [172].

6.3. Scalability
As AI models are being deployed in real-time and large-scale systems, scalability
has become a pressing challenge for XAI. Many existing explanation techniques, such as
Shapley and LIME values, can be computationally expensive, especially for models with a
large number of features or parameters [173]. For example, computing SHAP values [15] for
a deep learning model can involve extensive Monte Carlo simulations [174]. Such extensive
simulations undermine SHAP’s practical applicability for a real-time system with large
datasets [175]. Thus, increasing complexity of the underlying model results in rising cost
of the explanation. Recent works on more efficient explanation techniques, such as sparse
feature attribution methods and approximation techniques for Shapley values [168], have
addressed this scalability issue. However, the need for faster and scalable XAI methods
remains a major challenge, particularly in large-scale production environments where
time-sensitive decisions are critical.

6.4. Fairness and Privacy

Another significant challenge in XAI is to ensure that the explanation is free of bias
and discriminatory outcomes. AI models, trained with historical data, can inherit and even
magnify societal biases related to race, gender, or socioeconomic status [27]. Consequently,
XAI fairness is linked with data diversity [176], which refers to a model’s ability to represent
all types of objects in its output. For example, a content recommendation AI model should
recommend diverse contents to its users, rather than accurate yet similar contents [177].
To identify and mitigate this bias of AI models through explanation, the concept of fairness-
aware XAI has emerged as a critical area of research [135]. Counterfactual statements,
introduced by [56], can also be used to identify unfairness in the decision-making process
of an AI model. Moreover, if a biased model is explained in a way which makes it appear
fair, the risk of “explainability laundering” will arise [178]. Hence, it is imperative for XAI
to ensure fairness of AI models through an ethical approach.
XAI should also protect privacy while dealing with people’s personal data [179].
Explanation of an AI model has to be intelligible to the audience without compromising
privacy. Data governance, access protocols, and the quality and integrity of the data are
covered by privacy [180]. For example, a few images may be sufficient to violate a user’s
privacy even if the images are obfuscated [181]. Influential parameters of a neural network
can also be detected using some awkward input queries [182,183]. Cryptographic protocols
can be adopted by AI models to preserve individual and group privacy [184,185].

6.5. Human-Centric Interpretability

Human-centric interpretability is crucial to make an explanation intelligible to stake-
holders. According to cognitive psychology research, humans are not always adept at
interpreting the raw output and complex statistical metrics of a model [23]. Hence, an ex-
planation has to be tailored to the cognitive abilities of both domain experts and laypersons.
For instance, interactive and visualization-based explanation techniques may better fit the
mental model of a human user [28]. As the mental model of a human user is taken into
account to design explanations, XAI requires interdisciplinary collaboration among various
research disciplines, such as social science, politics, psychology, humanities, and law [186].
There needs to be a collaborative platform for XAI, where individual users can share how
data are collected, preprocessed, cleaned, modeled, and analyzed. Such collaboration will
enhance the accessibility of explanations provided using XAI methods.
Algorithms 2025, 18, 556 24 of 36

The trade-offs and challenges of XAI are complex and multifaceted. Future research
should be directed towards balancing these trade-offs, while addressing the practical
challenges of real-world deployment.

7. Future Scope of XAI

The future of XAI holds significant promise, driven by advancement in AI technologies,
increasing regulatory demands, and the growing need for transparency in high-stakes
decision-making. As AI systems become more integrated into our society, the demand for
interpretability and accountability will continue to rise. In this section, we describe the
scope of future research for XAI.

7.1. Formalization of Definition and Terminology

Future research should be focused on formalization of a systematic definition of
XAI [187]. A standard definition will facilitate the transfer of results among XAI platforms.
All synonymous but semantically different terms require to be standardized, so that the
same notions are not referred to using different names [187]. To address this persistent
lack of formalized terminology in XAI, the field should produce a compact taxonomy and
glossary to clearly distinguish core concepts, such as interpretability versus explainability,
local versus global explanation, and ante hoc versus post hoc explanation. Such distinction
will ensure understanding the same meaning of currently ambiguous terms across technical,
legal, and social science audiences [188]. Moreover, each term has to be paired with
an operational definition, and a precise mathematical or procedural specification. Such
specification should map each term to measurable properties, such as fidelity, stability,
and plausibility, leading to reproducible evaluation [189]. To report comparable XAI results
across domains, the XAI community should maintain a living ontology and benchmark
registry. This practice will enable researchers to test methods against the same task-and-
metric pairs [40].

7.2. Advancement in Explainability Techniques

One of the most exciting areas for the future of XAI lies in the development of more
sophisticated and effective explanation methods. Although current methods, such as
LIME [4] and SHAP [15], are promising, they are challenging to scale to large complex
models. Moreover, most of the existing explanation techniques cannot provide explanations
which are comprehensible across diverse application domains. Future research may focus
on developing explainability methods tailored to specific domains, such as healthcare,
finance, and criminal justice. Such customized explanation will provide context-specific
insights, which are crucial for decision-making. Furthermore, as deep learning continues to
evolve and becomes more complex, model-agnostic XAI approaches need to be improved
to provide consistent explanations [20].

7.3. Human-Centered AI and Cognitive Alignment

Future XAI research should emphasize human cognitive factors to design explanations
that are aligned with human understanding. Such alignment will reduce the cognitive
load of end-users and improve the trustworthiness of AI decisions [23]. For this purpose,
cognitive science and HCI will become integral parts of XAI research. Thus, explanations
can become more intelligible to a wide range of users, ranging from expert practitioners to
the general public.

7.4. Interactive Explanation

Another significant area for the future of XAI is the development of interactive ex-
planation interfaces [190], where users can actively engage with the model to explore its
Algorithms 2025, 18, 556 25 of 36

decision-making process. Interactive interfaces include features such as allowing users to

query a model, modify input, and assess the impact of different parameters on model out-
comes [30]. Such dynamic and adaptive explanations will provide a deeper understanding
of the model’s internal reasoning to users through feedback loops, resulting in improved
trustworthiness. Affective computing [191], where affordances are varied based on human
feedback, will play an important role to this effect.

7.5. Ethical XAI

As AI technologies continue to influence critical sectors, such as healthcare, finance,
and criminal justice, the need for ethical and fair AI models will become increasingly impor-
tant [192]. The future of XAI is likely to integrate ethical considerations with explainability
techniques, focusing on fairness, accountability, and transparency. Such an ethical approach
will enable XAI to identify decisions which may irrationally affect an individual or a group
of people [27].

7.6. Regulatory Compliance

The future of XAI will also be influenced by regulatory frameworks [193]. Various
governments and organizations are increasingly adopting laws and standards to ensure the
transparency of AI decisions. As regulations become stricter, new regulatory standards will
emerge for explainability in the future. Hence, XAI researchers need to collaborate with law-
makers to develop solutions which balance legal requirements with technical advancements.

7.7. Explainability for New AI Paradigms

As AI continues to evolve, new paradigms, such as federated learning [194] and re-
inforcement learning [195], are gaining prominence. The complexity of these models
introduces new challenges for interpretability. The future of XAI will involve the develop-
ment of explainability methods for these next-generation AI models [196]. For instance,
federated learning involves decentralized data processing. As data are not centrally located,
questions are raised regarding explanations provided of the decision-making process in
federated learning [197]. Similarly, reinforcement learning models, which learn through
interaction with their environment, will require new strategies for providing clear and
meaningful explanations of their decision-making.

7.8. Standardization of Evaluation Metrics

An important area for future research will be the standardization of evaluation metrics
for XAI. We show some key XAI evaluation metrics, along with their evaluation objectives
and benchmark datasets, in Table 9. Although several evaluation criteria, such as fidelity,
stability, and human trust currently exist [23], there is no universally accepted framework
for evaluating XAI methods. To address this, we propose a unified evaluation framework
to operationalize three foundational evaluation dimensions: fidelity, stability, and human
trust. Fidelity reflects how faithfully an explanation captures the underlying model’s
reasoning, stability shows how consistent the explanation is under input perturbations,
and human trust shows how effectively an explanation supports the user’s confidence.
We aggregate these three evaluation dimensions through a mathematical equation, as
shown below.
 
EXAI (x) = λ F w1 norm( R2 ) + w2 norm(Del) + w3 norm(CF)


+ λS w4 norm(ρ) + w5 norm(Topk ) + w6 1 − norm( L) (1)
 
+ λ T w7 norm(∆Perf) + w8 norm(∆Cal) + w9 norm(∆Det) + w10 norm( Tsubj )

λ F + λS + λ T = 1, λ F , λS , λ T ∈ [0, 1],
Algorithms 2025, 18, 556 26 of 36

w1 + w2 + w3 = 1, w4 + w5 + w6 = 1, w7 + · · · + w10 = 1, wi ∈ [0, 1].

We clarify the meaning of the symbols of this Equation (1) in Table 10. By default, each
evaluation dimension is equally important (λ F = λS = λ T = 13 ). However, based on the
priority of the stakeholder, these weights can be adjusted to reflect the relative importance of
the three evaluation dimensions in a given application domain. For example, safety-critical
applications may prioritize fidelity (highest weight for λ F ), while user-centric systems may
emphasize human trust (highest weight for λ T ). This evaluation scoring system, as shown
in Equation (1), can be applied to any XAI method to evaluate its explainability. Thus, our
proposed evaluation framework enables the structured evaluation and comparison of XAI
methods, leading to the development of more user-aligned solutions. Moreover, the future
of XAI will encompass the development of standardized evaluation protocols [21], which
can be used across different models, domains, and use cases. Such standardization will
enhance the comparison of various XAI methods in a quantitative manner.

Table 9. XAI evaluation metrics, objectives, and benchmark datasets.

Evaluation Metric Evaluation Objective Benchmark Dataset(s)

Faithfulness: does deleting/inserting top-ranked ImageNet, CIFAR-10 used in XAI
Deletion/Insertion [198]
features change the model output? saliency/deletion evaluation [199].
Local fidelity: how well a simple local surrogate “UCI Adult” used in LIME/SHAP analyses and
Local Surrogate R2 [198]
fits the black-box near an instance x? fairness/explainability studies [200].
Validity/plausibility: do counterfactuals flip
Counterfactuals [201] UCI Adult, German Credit, COMPAS [202].
the model?
Stability: is the feature ranking consistent under
Rank Correlation [190] CIFAR, ImageNet [199].
perturbations?
Stability: fraction of top-k features preserved
Top-k overlap [190] UCI Adult, COMPAS, ImageNet, PASCAL [198].
under perturbations.
Robustness: magnitude of explanation change per
Lipschitz Ratio [203] unit input change (the lower the change, ImageNet, CIFAR [203].
the better).
Human usefulness: does “Model + XAI” provide ERASER (NLP) and clinical datasets (MIMIC,
Task Performance [190] better human decision accuracy than CheXpert) used in human-centred XAI
“Model-only”? experiments [204].
Human calibration: does users’ confidence
Calibration (Brier) [190] ERASER, MIMIC, CheXpert [204].
improve after seeing explanation?
Appropriate reliance: do explanations help users MIMIC, CheXpert, ERASER, and controlled
Error Detection [205]
detect model errors? synthetic datasets [206].
Self-reported human trust (Likert scale) used Any human-study task (ERASER, MIMIC,
Subjective Trust [190]
alongside the behavioral metrics. ImageNet, UCI), along with questionnaires [206].

Table 10. Meanings of the symbols of Equation (1).

Symbol Meaning
EXAI (x) Evaluation score aggregating fidelity, stability, and human trust (range [0, 1] after normalization) for an instance x.
λ F , λS , λ T Weights of the three dimensions: fidelity, stability, and human trust, respectively; λ F + λS + λ T = 1.
w1 , . . . , w10 Internal sub-metric weights for the three evaluation dimensions.
norm(·) Min–max normalization operator mapping a raw metric to [0, 1].
R2 Local surrogate fidelity [198].
Del Deletion / insertion fidelity metric. Normalized Area Under Curve (AUC) drops when top features are removed [198].
CF Counterfactual validity [201].
ρ Spearman rank correlation between original and perturbed explanation rankings (stability) [207].
Topk Top-k overlap between important feature sets under perturbation [190].
L Lipschitz ratio for explanations [203] (magnitude of explanation change per unit input change).
Algorithms 2025, 18, 556 27 of 36

Table 10. Cont.

Symbol Meaning
∆Perf Change in human task performance (“Model + XAI” versus “Model-only”) [190].
∆Cal Improvement in calibration (Brier score) [190] because of explanation.
∆Det Improvement in error-detection rate (fraction of model errors users flag) [205].
Tsubj Aggregated subjective trust score (normalized Likert-based scale, mapped to [0, 1]) [190].

7.9. Economic and Sustainability Perspective

An economic perspective on XAI has attracted little attention to date in the literature,
yet this is no less important [52]. XAI can be a driving force to increase business value in
future. Explainability cost estimation, algorithmic propriety, and trade secrets are among
the factors which will inform the economic dimension of XAI in the future [52]. “Structural
Econometrics for Explainable AI” was proposed by [208] to highlight the connection
between AI and econometrics. To ensure sustainability, XAI products and services have to
be compliant with the United Nations’ Sustainable Development Goals (SDG) [209] for the
greater benefit of mankind.
The future scope of XAI is expansive. As AI continues to evolve and permeate new
domains, the importance of explainability will only grow. The future of XAI will be pivotal
in ensuring that AI systems are transparent, trustworthy, and aligned with human values.
Researchers, practitioners, and policymakers have to work collaboratively to seize the
opportunities presented by XAI for the future.

8. Conclusions
Explainable Artificial Intelligence (XAI) has emerged as a critical field in response
to the growing complexity and opacity of modern AI systems. As AI technologies are
increasingly deployed in high-stakes domains, such as healthcare, finance, law, and au-
tonomous systems, the demand for transparency, accountability, and trustworthiness has
intensified. This survey presented a comprehensive overview of XAI, including its founda-
tional definitions, key terminologies, taxonomy, application domains, trade-offs, and future
research directions. In this paper, we observed that the definition of explainability varies
significantly across stakeholders, ranging from a technical description for developers to
human-centric interpretation aligned with laypersons’ cognitive needs. Although the
terminologies of XAI remain somewhat fragmented, efforts to establish standardized vocab-
ularies and taxonomies will contribute to a more structured understanding of the XAI field.
The taxonomy of XAI techniques provided highlights the diversity and depth of technical
strategies to make models interpretable. Moreover, our review of application domains
demonstrated how explainability requirements differ based on contextual constraints, legal
requirements, and the potential consequences of AI decisions. Despite substantial progress,
XAI continues to face various challenges, such as the accuracy versus interpretability
trade-off, lack of a universal definition, and lack of standardized evaluation metrics. With
regard to the future scope of XAI, we focused on human-centered interactive explanation,
integrating explainability requirements with regulatory frameworks, and explainability for
novel AI paradigms, such as federated and reinforcement learning. Interdisciplinary collab-
oration among AI researchers, cognitive scientists, legal experts, and domain practitioners
is essential to align XAI with societal values.
To conclude, XAI represents not only a technical challenge, but also a sociotechnical
one. It intersects with issues of fairness, responsibility, and trust. As AI continues to shape
critical decisions and societal outcomes, the importance of effective, reliable, and accessible
explanations will only grow. Hence, investment in research, standards, and education
around explainability is essential for ensuring AI can be used for social good.
Algorithms 2025, 18, 556 28 of 36

Author Contributions: Conceptualization, M.S.H. and S.K.; methodology, M.S.H. and S.K.; soft-
ware, S.K.; validation, S.K., M.S.H. and K.A.; formal analysis, S.K. and M.S.H.; investigation,
S.K. and M.S.H.; resources, K.A.; data curation, K.A.; writing—original draft preparation, S.K.;
writing—review and editing, M.S.H.; visualization, S.K.; supervision, M.S.H. and K.A.; project ad-
ministration, K.A.; funding acquisition, K.A. All authors have read and agreed to the published
version of the manuscript.

Funding: This research was funded by VINNOVA (Sweden’s Innovation Agency) through the Digital
Stadsutveckling Campus Skellefteå project, grant number 2022-01188.

Data Availability Statement: No new data were created or analyzed in this study. Data sharing is
not applicable to this article.

Conflicts of Interest: The authors declare no conflicts of interest. The funders had no role in the design
of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or
in the decision to publish the results.

References
1. Walmsley, J. Artificial Intelligence and the Value of Transparency. AI Soc. 2021, 36, 585–595. [CrossRef]
2. Wiggerthale, J.; Reich, C. Explainable Machine Learning in Critical Decision Systems: Ensuring Safe Application and Correctness.
AI 2024, 5, 2864–2896. [CrossRef]
3. Emmert-Streib, F.; Yang, Z.; Feng, H.; Tripathi, S.; Dehmer, M. An Introductory Review of Deep Learning for Prediction Models
with Big Data. Front. Artif. Intell. 2020, 3, 4. [CrossRef]
4. Ribeiro, M.T.; Singh, S.; Guestrin, C. Why Should I Trust You? Explaining the Predictions of Any Classifier. In Proceedings of the
22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, 13–17 August
2016; pp. 1135–1144. [CrossRef]
5. Gunning, D. Explainable Artificial Intelligence (XAI); DARPA: Arlington, VA, USA, 2017. Available online: https://www.darpa.
mil/program/explainable-artificial-intelligence (accessed on 6 June 2025).
6. Gunning, D.; Aha, D.W. DARPA’s Explainable Artificial Intelligence (XAI) Program. AI Mag. 2019, 40, 44–58.
7. Mahto, M.K. Explainable Artificial Intelligence: Fundamentals, Approaches, Challenges, XAI Evaluation, and Validation. In
Explainable Artificial Intelligence for Autonomous Vehicles; CRC Press: Boca Raton, FL, USA, 2025; pp. 25–49.
8. Doshi-Velez, F.; Kim, B. Towards a rigorous science of interpretable machine learning. arXiv 2017, arXiv:1702.08608. [CrossRef]
9. Gilpin, L.H.; Bau, D.; Zoran, D.; Pezzotti, N.; Bach, S.; Wexler, J. Explaining Explanations: An Overview of Interpretability of
Machine Learning. In Proceedings of the 2018 ICML Workshop on Human Interpretability in Machine Learning (WHI 2018),
Stockholm, Sweden, 15 July 2018. [CrossRef]
10. Li, X.; Xiong, H.; Li, X.; Wu, X.; Zhang, X.; Liu, J.; Dou, D. Interpretable Deep Learning: Interpretation, Interpretability,
Trustworthiness, and Beyond. Knowl. Inf. Syst. 2022, 64, 3197–3234. [CrossRef]
11. Albahri, A.S.; Duhaim, A.M.; Fadhel, M.A.; Alnoor, A.; Baqer, N.S.; Alzubaidi, L.; Deveci, M. A Systematic Review of Trustworthy
and Explainable Artificial Intelligence in Healthcare: Assessment of Quality, Bias Risk, and Data Fusion. Inf. Fusion 2023, 96,
156–191. [CrossRef]
12. Caruana, R.; Gehrke, J.; Koch, C.; Sturm, M. Intelligible Models for Healthcare: Predicting Pneumonia Risk and Hospital 30-Day
Readmission. In Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Sydney,
NSW, Australia, 10–13 August 2015; Volume 21, pp. 1721–1730. [CrossRef]
13. Zhang, Q.; Wu, Y.N.; Zhu, S.C. Interpretable Convolutional Neural Networks. In Proceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition (CVPR), Salt Lake City, UT, USA, 18–22 June 2018; pp. 8827–8836. [CrossRef]
14. Silver, D.; Schrittwieser, J.; Simonyan, K.; Antonoglou, I.; Huang, A.; Guez, A.; Hubert, T.; Baker, L.; Lai, M.; Bolton, A.; et al.
Mastering the Game of Go without Human Knowledge. Nature 2017, 550, 354–359. [CrossRef] [PubMed]
15. Lundberg, S.M.; Lee, S.I. A Unified Approach to Interpreting Model Predictions. In Proceedings of the Advances in Neural
Information Processing Systems, Long Beach, CA, USA, 4–9 December 2017; pp. 4765–4774. Available online: https://arxiv.org/
abs/1705.07874 (accessed on 28 August 2025).
16. Khan, A.; Ali, A.; Khan, J.; Ullah, F.; Faheem, M. Exploring Consistent Feature Selection for Software Fault Prediction: An
XAI-based model-agnostic Approach. IEEE Access 2025, 13, 75493–75524. . [CrossRef]
17. Assis, A.; Dantas, J.; Andrade, E. The Performance-Interpretability Trade-Off: A Comparative Study of Machine Learning Models.
J. Reliab. Intell. Environ. 2025, 11, 1. [CrossRef]
18. Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning; MIT Press: Cambridge, MA, USA, 2016.
Algorithms 2025, 18, 556 29 of 36

19. Serradilla, O.; Zugasti, E.; Rodriguez, J.; Zurutuza, U. Deep Learning Models for Predictive Maintenance: A Survey, Comparison,
Challenges and Prospects. Appl. Intell. 2022, 52, 10934–10964. [CrossRef]
20. Lipton, Z.C. The Mythos of Model Interpretability. Commun. ACM 2016, 59, 36–43. [CrossRef]
21. Lopes, P.; Silva, E.; Braga, C.; Oliveira, T.; Rosado, L. XAI Systems Evaluation: A Review of Human and Computer-Centred
Methods. Appl. Sci. 2022, 12, 9423. [CrossRef]
22. Kadir, M.A.; Mosavi, A.; Sonntag, D. Evaluation Metrics for XAI: A Review, Taxonomy, and Practical Applications. In Proceedings
of the 2023 IEEE 27th International Conference on Intelligent Engineering Systems (INES), Tihany, Hungary, 19–21 July 2023;
pp. 000111–000124.
23. Miller, T. Explanation in Artificial Intelligence: Insights from the Social Sciences. Artif. Intell. 2019, 267, 1–38. [CrossRef]
24. Abdul, A.; Vermeulen, J.; Wang, D.; Lim, B.Y.; Kankanhalli, M. Trends and Trajectories for Explainable, Accountable and
Intelligible Systems: An HCI Research Agenda. In Proceedings of the 2018 CHI Conference on Human Factors in Computing
Systems, Montreal, QC, Canada, 21–26 April 2018; pp. 1–18. [CrossRef]
25. Kabir, S.; Hossain, M.S.; Andersson, K. An Advanced Explainable Belief Rule-Based Framework to Predict the Energy Consump-
tion of Buildings. Energies 2024, 17, 1797. [CrossRef]
26. Kabir, S.; Hossain, M.S.; Andersson, K. A Semi-Supervised-Learning-Aided Explainable Belief Rule-Based Approach to Predict
the Energy Consumption of Buildings. Algorithms 2025, 18, 305. [CrossRef]
27. Barocas, S.; Hardt, M.; Narayanan, A. Fairness and Machine Learning; 2019. Available online: https://fairmlbook.org/ (accessed
on 9 June 2025).
28. Kulesza, T.; Stumpf, S.; Burnett, M.; Wong, W.K.; Striegel, A. Principles of Explanatory Debugging to Assist Machine Learning. In
Proceedings of the 2013 CHI Conference on Human Factors in Computing Systems (CHI ’13), Paris, France, 27 April–2 May 2013;
pp. 1519–1528. [CrossRef]
29. Zhou, B.; Khosla, A.; Lapedriza, A.; Oliva, A.; Torralba, A. Learning Deep Features for Discriminative Localization. In Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, NV, USA, 27–30 June 2016; pp. 2921–2929.
[CrossRef]
30. Stumpf, S.; Rajaram, V.; Li, L.; Burnett, M.; Dietterich, T.G.; Sullivan, E.; Drummond, R. Interacting with Machine Learning
Systems: The Case for Explanation and Beyond. In Proceedings of the 14th International Conference on Intelligent User Interfaces
(IUI ’09), Sanibel Island, FL, USA, 8–11 February 2009; pp. 109–118.
31. Calders, T.; Verwer, S. Three Naive Bayes Approaches for Discrimination-Free Classification. Data Min. Knowl. Discov. 2010, 21,
277–292. [CrossRef]
32. Goodman, B.; Flaxman, S. European Union Regulations on Algorithmic Decision-Making and a “Right to Explanation”. Proc.
Mach. Learn. Res. 2017, 81, 1–10. [CrossRef]
33. Moral, A.; Castiello, C.; Magdalena, L.; Mencar, C. Explainable Fuzzy Systems; Springer International Publishing: New York, NY,
USA, 2021. [CrossRef]
34. Shortliffe, E. (Ed.) Computer-Based Medical Consultations: MYCIN; Elsevier: Amsterdam, The Netherlands, 2012; Volume 2.
35. Shortliffe, E.H.; Buchanan, B.G. A Model of Inexact Reasoning in Medicine. Math. Biosci. 1975, 23, 351–379. [CrossRef]
36. Wang, W.; Yang, Y.; Wu, F. Towards Data- and Knowledge-Driven AI: A Survey on Neuro-Symbolic Computing. IEEE Trans.
Pattern Anal. Mach. Intell. 2024, 47, 878–899. [CrossRef]
37. Liang, B.; Wang, Y.; Tong, C. AI Reasoning in Deep Learning Era: From Symbolic AI to Neural–Symbolic AI. Mathematics 2025,
13, 1707. [CrossRef]
38. Fernandez-Quilez, A. Deep Learning in Radiology: Ethics of Data and on the Value of Algorithm Transparency, Interpretability
and Explainability. AI Ethics 2023, 3, 257–265. [CrossRef]
39. Barocas, S.; Friedler, S.; Hardt, M.; Kroll, J.; Venkatasubramanian, S.; Wallach, H. The FAT-ML Workshop Series on Fairness,
Accountability, and Transparency in Machine Learning. Available online: http://www.fatml.org/ (accessed on 9 June 2025).
40. Longo, L.; Brcic, M.; Cabitza, F.; Choi, J.; Confalonieri, R.; Del Ser, J.; Stumpf, S. Explainable Artificial Intelligence (XAI) 2.0: A
Manifesto of Open Challenges and Interdisciplinary Research Directions. Inf. Fusion 2024, 106, 102301. [CrossRef]
41. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. ImageNet Classification with Deep Convolutional Neural Networks. Adv. Neural Inf.
Process. Syst. 2012, 25, 1097–1105. [CrossRef]
42. Souza, V.F.; Cicalese, F.; Laber, E.; Molinaro, M. Decision Trees with Short Explainable Rules. Adv. Neural Inf. Process. Syst. 2022,
35, 12365–12379. [CrossRef]
43. Agarwal, C.; Krishna, S.; Saxena, E.; Pawelczyk, M.; Johnson, N.; Puri, I.; Lakkaraju, H. OpenXAI: Towards a Transparent
Evaluation of Model Explanations. Adv. Neural Inf. Process. Syst. 2022, 35, 15784–15799.
44. Basti, G.; Vitiello, G. Deep Learning Opacity, and the Ethical Accountability of AI Systems: A New Perspective. In The Logic of
Social Practices II; Springer Nature Switzerland: Cham, Switzerland, 2023; pp. 21–73. [CrossRef]
45. Alvarez-Melis, D.; Jaakkola, T.S. On the Robustness of Interpretability Methods. arXiv 2018, arXiv:1806.08049. [CrossRef]
Algorithms 2025, 18, 556 30 of 36

46. Binns, R. Fairness in Machine Learning: Lessons from Political Philosophy. In Proceedings of the 2018 Conference on Fairness,
Accountability and Transparency (FAT), New York, NY, USA, 23–24 February 2018; pp. 149–159.
47. Retzlaff, C.O.; Angerschmid, A.; Saranti, A.; Schneeberger, D.; Roettger, R.; Mueller, H.; Holzinger, A. Post-Hoc vs Ante-Hoc
Explanations: XAI Design Guidelines for Data Scientists. Cogn. Syst. Res. 2024, 86, 101243. [CrossRef]
48. Rudin, C.; Chen, C.; Chen, Z.; Huang, H.; Semenova, L.; Zhong, C. Interpretable Machine Learning: Fundamental Principles and
10 Grand Challenges. Stat. Surv. 2022, 16, 1–85. [CrossRef]
49. Necula, S.-C.; Dumitriu, F.; Greavu-S, erban, V. A Systematic Literature Review on Using Natural Language Processing in Software
Requirements Engineering. Electronics 2024, 13, 2055. [CrossRef]
50. Page, M.J.; McKenzie, J.E.; Bossuyt, P.M.; Boutron, I.; Hoffmann, T.C.; Mulrow, C.D.; Moher, D. The PRISMA 2020 statement: An
updated guideline for reporting systematic reviews. Int. J. Surg. 2021, 88, 105906. [CrossRef]
51. Singh, V.K.; Singh, P.; Karmakar, M.; Leta, J.; Mayr, P. The journal coverage of Web of Science, Scopus and Dimensions: A
comparative analysis. Scientometrics 2021, 126, 5113–5142. [CrossRef]
52. Adadi, A.; Berrada, M. Peeking Inside the Black-Box: A Survey on Explainable Artificial Intelligence (XAI). IEEE Access 2018, 6,
52138–52160. [CrossRef]
53. Guidotti, R.; Monreale, A.; Ruggieri, S.; Turini, F.; Giannotti, F.; Pedreschi, D. A Survey of Methods for Explaining Black Box
Models. ACM Comput. Surv. 2018, 51, 1–42. [CrossRef]
54. Rudin, C. Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and Use Interpretable Models Instead.
Nat. Mach. Intell. 2019, 1, 206–215. [CrossRef]
55. Szepannek, G.; Lübke, K. How Much Do We See? On the Explainability of Partial Dependence Plots for Credit Risk Scoring.
Argum. Oeconomica 2023, 1, 137–150. [CrossRef]
56. Wachter, S.; Mittelstadt, B.; Russell, C. Counterfactual Explanations without Opening the Black Box: Automated Decisions and
the GDPR. Harv. J. Law Technol. 2017, 31, 841. [CrossRef]
57. Ribeiro, M.T.; Singh, S.; Guestrin, C. Anchors: High-Precision Model-Agnostic Explanations. Proc. AAAI Conf. Artif. Intell.
2018, 32. [CrossRef]
58. Goldstein, A.; Kapelner, A.; Bleich, J.; Pitkin, E. Predicting Consumer Choices from Complex Behavioral Data. In Proceedings of
the 21st ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Sydney, Australia, 10–13 August
2015; pp. 1315–1324.
59. Liu, Z.; Zhang, J.; Wang, Y. Interpretable Machine Learning for Predicting Sepsis Risk in ICU Patients. Sci. Rep. 2025, 15, 12345.
[CrossRef]
60. Hariharan, S.; Rejimol Robinson, R.R.; Prasad, R.R.; Thomas, C.; Balakrishnan, N. XAI for Intrusion Detection System: Comparing
Explanations Based on Global and Local Scope. J. Comput. Virol. Hack. Tech. 2023, 19, 217–239. [CrossRef]
61. Ponce-Bobadilla, A.V.; Schmitt, V.; Maier, C.S.; Mensing, S.; Stodtmann, S. Practical guide to SHAP analysis: Explaining supervised
machine learning model predictions in drug development. Clin. Transl. Sci. 2024, 17, e70056. [CrossRef]
62. Ruiz de San Martín, R.; Morales-Hernández, C.; Barberá, C.; Martínez-Cortés, C.; Banegas-Luna, A.J.; Segura-Méndez, F.J.;
Pérez-Sánchez, H.; Morales-Moreno, I.; Hernández-Morante, J.J. Global and Local Interpretable Machine Learning Allow Early
Prediction of Unscheduled Hospital Readmission. Mach. Learn. Knowl. Extr. 2024, 6, 1653–1666. [CrossRef]
63. Sathyan, A.; Weinberg, A.I.; Cohen, K. Interpretable AI for Bio-Medical Applications. Complex Eng. Syst. 2022, 2, 18. [CrossRef]
64. Wang, Y. A Comparative Analysis of Model Agnostic Techniques for Explainable Artificial Intelligence. Res. Rep. Comput. Sci.
2024, 3, 25–33. [CrossRef]
65. Selvaraju, R.R.; Cogswell, M.; Das, A.; Vedantam, R.; Parikh, D.; Batra, D. Grad-CAM: Visual Explanations from Deep Networks
via Gradient-Based Localization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), Venice, Italy,
22–29 October 2017; pp. 618–626. [CrossRef]
66. Bifarin, O.O. Interpretable Machine Learning with Tree-Based Shapley Additive Explanations: Application to Metabolomics
Datasets for Binary Classification. PLoS ONE 2023, 18, e0284315. [CrossRef]
67. Molnar, C. Interpretable Machine Learning: A Guide for Making Black Box Models Explainable. 2022. Available online:
https://christophm.github.io/interpretable-ml-book/ (accessed on 10 June 2025).
68. Feng, J.; Liang, J.; Qiang, Z.; Hao, Y.; Li, X.; Li, L.; Wei, H. A Hybrid Stacked Ensemble and Kernel SHAP-Based Model for
Intelligent Cardiotocography Classification and Interpretability. BMC Med. Inform. Decis. Mak. 2023, 23, 273. [CrossRef]
69. Vilone, G.; Longo, L. Classification of Explainable Artificial Intelligence Methods through Their Output Formats. Mach. Learn.
Knowl. Extract. 2021, 3, 615–661. [CrossRef]
70. Cremades, A.; Hoyas, S.; Vinuesa, R. Additive-Feature-Attribution Methods: A Review on Explainable Artificial Intelligence for
Fluid Dynamics and Heat Transfer. Int. J. Heat Fluid Flow 2025, 112, 109662. [CrossRef]
71. Bauer, K.; von Zahn, M.; Hinz, O. Expl(AI)ned: The Impact of Explainable Artificial Intelligence on Users’ Information Processing.
Inf. Syst. Res. 2023, 34, 1582–1602. [CrossRef]
Algorithms 2025, 18, 556 31 of 36

72. Mertes, S.; Huber, T.; Weitz, K.; Heimerl, A.; André, E. GANterfactual—Counterfactual Explanations for Medical Non-Experts
Using Generative Adversarial Learning. Front. Artif. Intell. 2022, 5, 825565. [CrossRef] [PubMed]
73. Simonyan, K.; Vedaldi, A.; Zisserman, A. Deep Inside Convolutional Networks: Visualising Image Classification Models and
Saliency Maps. arXiv 2014, arXiv:1312.6034. [CrossRef]
74. Saporta, A.; Gui, X.; Agrawal, A.; Pareek, A.; Truong, S.Q.H.; Nguyen, C.D.T.; Ngo, V.-D.; Seekins, J.; Blankenberg, F.G.; Ng, A.Y.;
et al. Benchmarking Saliency Methods for Chest X-ray Interpretation. Nat. Mach. Intell. 2022, 4, 867–878. [CrossRef]
75. Ford, C.; Keane, M.T. Explaining Classifications to Non-Experts: An XAI User Study of Post-Hoc Explanations for a Classifier
When People Lack Expertise. In Proceedings of the International Conference on Pattern Recognition, Cham, Switzerland, 21–25
August 2022; Springer Nature Switzerland: Cham, Switzerland, 2022; pp. 246–260. [CrossRef]
76. Bashir, Z.; Lin, M.; Feragen, A.; Mikolaj, K.; Taksøe-Vester, C.; Christensen, A.N.; Svendsen, M.B.S.; Fabricius, M.H.; Andreasen,
L.; Nielsen, M.; et al. Clinical Validation of Explainable AI for Fetal Growth Scans through Multi-Level, Cross-Institutional
Prospective End-User Evaluation. Sci. Rep. 2025, 15, 2074. [CrossRef]
77. Slack, D.; Krishna, S.; Lakkaraju, H.; Singh, S. Explaining Machine Learning Models with Interactive Natural Language
Conversations Using TalkToModel. Nat. Mach. Intell. 2023, 5, 873–883. [CrossRef]
78. Bertrand, A.; Viard, T.; Belloum, R.; Eagan, J.R.; Maxwell, W. On Selective, Mutable and Dialogic XAI: A Review of What Users Say
about Different Types of Interactive Explanations. In Proceedings of the 2023 CHI Conference on Human Factors in Computing
Systems, Hamburg, Germany, 23–28 April 2023; pp. 1–21. [CrossRef]
79. Nachtigall, M.; Do, L.N.Q.; Bodden, E. Explaining Static Analysis—A Perspective. In Proceedings of the 2019 34th IEEE/ACM
International Conference on Automated Software Engineering Workshop (ASEW), San Diego, CA, USA, 11–15 November 2019;
IEEE: Piscataway, NJ, USA, 2019; pp. 29–32. [CrossRef]
80. Shi, H.; Yang, N.; Yang, X.; Tang, H. Clarifying Relationship between PM2.5 Concentrations and Spatiotemporal Predictors Using
Multi-Way Partial Dependence Plots. Remote Sens. 2023, 15, 358. [CrossRef]
81. Amershi, S.; Chickering, M.; Drucker, S.M.; Lee, B.; Simard, P.; Suh, J. ModelTracker: Redesigning Performance Analysis Tools for
Machine Learning. In Proceedings of the CHI ’14: SIGCHI Conference on Human Factors in Computing Systems, Toronto, ON,
Canada, 26 April–1 May 2014; pp. 337–346. [CrossRef]
82. Metsch, J.M.; Saranti, A.; Angerschmid, A.; Pfeifer, B.; Klemt, V.; Holzinger, A.; Hauschild, A.-C. CLARUS: An Interactive
Explainable AI Platform for Manual Counterfactuals in Graph Neural Networks. J. Biomed. Inform. 2024, 150, 104600. [CrossRef]
[PubMed]
83. Delaney, E.; Pakrashi, A.; Greene, D.; Keane, M.T. Counterfactual Explanations for Misclassified Images: How Human and
Machine Explanations Differ. Artif. Intell. 2023, 324, 103995. [CrossRef]
84. Hassan, S.U.; Abdulkadir, S.J.; Zahid, M.S.M.; Al-Selwi, S.M. Local Interpretable Model-Agnostic Explanation Approach for
Medical Imaging Analysis: A Systematic Literature Review. Comput. Biol. Med. 2025, 185, 109569. [CrossRef]
85. Hwang, H.; Bell, A.; Fonseca, J.; Pliatsika, V.; Stoyanovich, J.; Whang, S.E. SHAP-based Explanations are Sensitive to Feature
Representation. In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, Athens, Greece,
23–26 June 2025; pp. 1588–1601. [CrossRef]
86. Sarti, G.; Feldhus, N.; Qi, J.; Nissim, M.; Bisazza, A. Democratizing Advanced Attribution Analyses of Generative Language
Models with the Inseq Toolkit. In Proceedings of the Joint of the 2nd World Conference on eXplainable Artificial Intelligence
Late-Breaking Work, Demos and Doctoral Consortium, xAI-2024: LB/D/DC, Valletta, Malta, 17–19 July 2024; pp. 289–296.
87. Miglani, V.; Yang, A.; Markosyan, A.H.; Garcia-Olano, D.; Kokhlikyan, N. Using Captum to Explain Generative Language Models.
In Proceedings of the 5th Workshop for Natural Language Processing Open Source Software (NLP-OSS), Singapore, 6 December
2023; Association for Computational Linguistics: Singapore, 2023; pp. 165–173. [CrossRef]
88. Simuni, G. Explainable AI in ML: The Path to Transparency and Accountability. Int. J. Recent Adv. Multidiscip. Res. 2024, 11,
10531–10536.
89. Ganapavarapu, G.; Mukherjee, S.; Martinez Gil, N.; Sarpatwar, K.; Rajasekharan, A.; Dhurandhar, A.; Arya, V.; Vaculin, R. AI
Explainability 360 Toolkit for Time-Series and Industrial Use Cases. In Proceedings of the 29th ACM SIGKDD Conference on
Knowledge Discovery and Data Mining, 6–10 August 2023; pp. 5777–5778. [CrossRef]
90. Rajkomar, A.; Hardt, M.; Howell, M.D.; Corrado, G.; Chin, M.H. Ensuring Fairness in Machine Learning for Healthcare. Lancet
2019, 393, 110–111.
91. Wang, H.; Wang, Z.; Du, M.; Yang, F.; Zhang, Z.; Ding, S.; Mardziel, P.; Hu, X. Score-CAM: Score-Weighted Visual Explanations
for Convolutional Neural Networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition
Workshops (CVPRW), Seattle, WA, USA, 14–19 June 2020; pp. 111–119. [CrossRef]
92. Musthafa M, M.; Mahesh, T.R.; Kumar V, V.; Guluwadi, S. Enhancing Brain Tumor Detection in MRI Images through Explainable
AI Using Grad-CAM with ResNet 50. BMC Med. Imaging 2024, 24, 107. [CrossRef]
Algorithms 2025, 18, 556 32 of 36

93. Alkhanbouli, R.; Matar Abdulla Almadhaani, H.; Alhosani, F.; Simsekler, M.C.E. The Role of Explainable Artificial Intelligence in
Disease Prediction: A Systematic Literature Review and Future Research Directions. BMC Med. Inform. Decis. Mak. 2025, 25, 110.
[CrossRef]
94. Chaddad, A.; Peng, J.; Xu, J.; Bouridane, A. Survey of Explainable AI Techniques in Healthcare. Sensors 2023, 23, 634. [CrossRef]
[PubMed]
95. Vanitha, K.; Mahesh, T.R.; Kumar, V.V.; Guluwadi, S. Enhanced Tuberculosis Detection Using Vision Transformers and Explainable
AI with a Grad-CAM Approach on Chest X-rays. BMC Med. Imaging 2025, 25, 96. [CrossRef]
96. Lu, S.; Chen, R.; Wei, W.; Belovsky, M.; Lu, X. Understanding Heart Failure Patients EHR Clinical Features via SHAP Interpretation
of Tree-Based Machine Learning Model Predictions. In AMIA Annual Symposium Proceedings; AMIA: San Francisco, CA, USA,
2022; Volume 2021, p. 813.
97. Goethals, S.; Martens, D.; Evgeniou, T. Manipulation risks in explainable AI: The implications of the disagreement problem.
In Joint European Conference on Machine Learning and Knowledge Discovery in Databases; Springer Nature Switzerland: Cham,
Switzerland, 2023; pp. 185–200. [CrossRef]
98. Mishra, A.K.; Tyagi, A.K.; Richa; Patra, S.R. Introduction to Machine Learning and Artificial Intelligence in Banking and
Finance. In Applications of Block Chain Technology and Artificial Intelligence: Lead-ins in Banking, Finance, and Capital Market; Springer
International Publishing: Cham, Switzerland, 2024; pp. 239–290. [CrossRef]
99. Cao, L. AI in Finance: Challenges, Techniques, and Opportunities. ACM Comput. Surv. 2022, 55, 1–38. [CrossRef]
100. Akkalkot, A.I.; Kulshrestha, N.; Sharma, G.; Sidhu, K.S.; Palimkar, S.S. Challenges and Opportunities in Deploying Explainable
AI for Financial Risk Assessment. In Proceedings of the 2025 International Conference on Pervasive Computational Technologies
(ICPCT), Pune, India, 8–9 February 2025; pp. 382–386. [CrossRef]
101. De Lange, P.E.; Melsom, B.; Vennerød, C.B.; Westgaard, S. Explainable AI for credit assessment in banks. J. Risk Financ. Manag.
2022, 15, 556. [CrossRef]
102. Freeborough, W.; van Zyl, T. Investigating Explainability Methods in Recurrent Neural Network Architectures for Financial Time
Series Data. Appl. Sci. 2022, 12, 1427. [CrossRef]
103. Černevičienė, J.; Kabašinskas, A. Explainable Artificial Intelligence (XAI) in Finance: A Systematic Literature Review. Artif. Intell.
Rev. 2024, 57, 216. [CrossRef]
104. Zhao, Z.; Bai, T. Financial Fraud Detection and Prediction in Listed Companies Using SMOTE and Machine Learning Algorithms.
Entropy 2022, 24, 1157. [CrossRef] [PubMed]
105. Sonkavde, G.; Dharrao, D.S.; Bongale, A.M.; Deokate, S.T.; Doreswamy, D.; Bhat, S.K. Forecasting Stock Market Prices Using
Machine Learning and Deep Learning Models: A Systematic Review, Performance Analysis and Discussion of Implications. Int. J.
Financ. Stud. 2023, 11, 94. [CrossRef]
106. Li, Y.; Stasinakis, C.; Yeo, W.M. A Hybrid XGBoost-MLP Model for Credit Risk Assessment on Digital Supply Chain Finance.
Forecasting 2022, 4, 184–207. [CrossRef]
107. Choudhary, H.; Orra, A.; Sahoo, K.; Thakur, M. Risk-Adjusted Deep Reinforcement Learning for Portfolio Optimization: A
Multi-Reward Approach. Int. J. Comput. Intell. Syst. 2025, 18, 126. [CrossRef]
108. Zhou, Y.; Li, H.; Xiao, Z.; Qiu, J. A User-Centered Explainable Artificial Intelligence Approach for Financial Fraud Detection.
Financ. Res. Lett. 2023, 58, 104309. [CrossRef]
109. Khan, F.S.; Mazhar, S.S.; Mazhar, K.; AlSaleh, D.A.; Mazhar, A. Model-Agnostic Explainable Artificial Intelligence Methods in
Finance: A Systematic Review, Recent Developments, Limitations, Challenges and Future Directions. Artif. Intell. Rev. 2025,
58, 232. [CrossRef]
110. Christin, A. Predictive Algorithms and Criminal Sentencing. In The Decisionist Imagination: Sovereignty, Social Science and Democracy
in the 20th Century; Steinmetz, G., Berkowitz, R., Eds.; Berghahn Books: New York, NY, USA, 2018; pp. 272–294. [CrossRef]
111. Raji, I.D.; Buolamwini, J. Actionable Auditing: Investigating the Impact of Publicly Naming Biased Performance Results of
Commercial AI Products. In Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society (AIES ’19), Honolulu, HI,
USA, 27–28 January 2019; ACM: New York, NY, USA, 2019; pp. 429–435. [CrossRef]
112. Mu, D.; Zhang, S.; Zhu, T.; Zhou, Y.; Zhang, W. Prediction of Recidivism and Detection of Risk Factors under Different Time
Windows Using Machine Learning Techniques. Soc. Sci. Comput. Rev. 2024, 42, 1379–1402. [CrossRef]
113. Zhang, X.; Liu, L.; Lan, M.; Song, G.; Xiao, L.; Chen, J. Interpretable Machine Learning Models for Crime Prediction. Comput.
Environ. Urban Syst. 2022, 94, 101789. [CrossRef]
114. Arayola, M.M.; Tal, I.; Connolly, R.; Saber, T.; Bendechache, M. Ethics and Trustworthiness of AI for Predicting the Risk of
Recidivism: A Systematic Literature Review. Information 2023, 14, 426. [CrossRef]
115. Eren, O.; Owens, E. Economic Booms and Recidivism. J. Quant. Criminol. 2024, 40, 343–372. [CrossRef]
116. Leung, C.; Mehrnia, B. A Multi-Model Approach to Legal Judgment Prediction Using Advanced Knowledge Integration
Techniques. J. Comput. Sci. Artif. Intell. 2024, 1, 9–16. [CrossRef]
Algorithms 2025, 18, 556 33 of 36

117. Lee, Y.; O, S.; Eck, J.E. Improving Recidivism Forecasting with a Relaxed Naïve Bayes Classifier. Crime Delinq. 2023, 71, 89–117.
(Original work published 2025). [CrossRef]
118. Cesario, E.; Lindia, P.; Vinci, A. Multi-Density Crime Predictor: An Approach to Forecast Criminal Activities in Multi-Density
Crime Hotspots. J. Big Data 2024, 11, 75. [CrossRef]
119. Pires, R.; de Souza, F.C.; Rosa, G.; Lotufo, R.; Nogueira, R. Sequence-to-Sequence Models for Extracting Information from
Registration and Legal Documents. In Document Analysis Systems. DAS 2022; Uchida, S., Barney, E., Eglin, V., Eds.; Lecture Notes
in Computer Science, Vol. 13237; Springer: Cham, Switzerland, 2022; pp. 73–88. [CrossRef]
120. Shamoo, Y. The Role of Explainable AI (XAI) in Forensic Investigations. In Digital Forensics in the Age of AI; IGI Global Scientific
Publishing: Cham, Switzerland, 2024; pp. 31–62. [CrossRef]
121. Fine, A.; Marsh, S. Judicial Leadership Matters (Yet Again): The Association between Judge and Public Trust for Artificial
Intelligence in Courts. Discov. Artif. Intell. 2024, 4, 44. [CrossRef]
122. Erdoğanyılmaz, C. A New Explainable AI Approach to Legal Judgement Prediction: Detecting Model Uncertainty and Analyzing
the Alignment Between Judges and Models. In Proceedings of the 2024 Innovations in Intelligent Systems and Applications
Conference (ASYU), Ankara, Turkey, 16–18 October 2024; pp. 1–6. [CrossRef]
123. Zafar, A. Artificial Intelligence in Autonomous Systems: Challenges and Opportunities. Res. Corrid. J. Eng. Sci. 2024, 1, 182–193.
124. Goodall, N.J. Machine Ethics and Automated Vehicles. In Road Vehicle Automation; Meyer, G., Beiker, S., Eds.; Lecture Notes in
Mobility; Springer International Publishing: Cham, Switzerland, 2014; pp. 93–102. [CrossRef]
125. Gartner, J. Autonomy and Accountability in Autonomous Systems. IEEE Intell. Syst. 2017, 32, 8–12.
126. Kuznietsov, A.; Gyevnar, B.; Wang, C.; Peters, S.; Albrecht, S.V. Explainable AI for Safe and Trustworthy Autonomous Driving: A
Systematic Review. IEEE Trans. Intell. Transp. Syst. 2024, 25, 19342–19364. [CrossRef]
127. Nazat, S.; Li, L.; Abdallah, M. XAI-ADS: An Explainable Artificial Intelligence Framework for Enhancing Anomaly Detection in
Autonomous Driving Systems. IEEE Access 2024, 12, 48583–48607. [CrossRef]
128. Yang, W.; Wei, Y.; Wei, H.; Chen, Y.; Huang, G.; Li, X.; Li, R.; Yao, N.; Wang, X.; Gu, X.; et al. Survey on Explainable AI: From
Approaches, Limitations and Applications Aspects. Hum.-Cent. Intell. Syst. 2023, 3, 161–188. [CrossRef]
129. Dong, J.; Chen, S.; Miralinaghi, M.; Chen, T.; Li, P.; Labi, S. Why did the AI make that decision? Towards an explainable artificial
intelligence (XAI) for autonomous driving systems. Transp. Res. Part C Emerg. Technol. 2023, 156, 104358. [CrossRef]
130. Hauptman, A.I.; Schelble, B.G.; Duan, W.; Flathmann, C.; McNeese, N.J. Understanding the influence of AI autonomy on AI
explainability levels in human-AI teams using a mixed methods approach. Cogn. Technol. Work 2024, 26, 435–455. [CrossRef]
131. Xu, Y.; Wu, D.; Zhou, M.; Yang, J. Deep learning-based point cloud classification of obstacles for intelligent vehicles. World Electr.
Veh. J. 2025, 16, 80. [CrossRef]
132. Wang, T.; Fu, Y.; Cheng, X.; Li, L.; He, Z.; Xiao, Y. Vehicle trajectory prediction algorithm based on hybrid prediction model with
multiple influencing factors. Sensors 2025, 25, 1024. [CrossRef]
133. Javaid, S.; Khan, M.A.; Fahim, H.; He, B.; Saeed, N. Explainable AI and monocular vision for enhanced UAV navigation in smart
cities: Prospects and challenges. Front. Sustain. Cities 2025, 7, 1561404. [CrossRef]
134. Rahevar, M.; Darji, S. The Adoption of AI-Driven Chatbots into a Recommendation for E-Commerce Systems to Targeted
Customer in the Selection of Product. Int. J. Manag. Econ. Commer. 2024, 1, 128–137. [CrossRef]
135. Dastin, J. Amazon Scraps Secret AI Recruiting Tool That Showed Bias Against Women. Reuters, 10 October 2018. Available
online: https://www.taylorfrancis.com/chapters/edit/10.1201/9781003278290-44/amazon-scraps-secret-ai-recruiting-tool-
showed-bias-women-jeffrey-dastin (accessed on 28 August 2025).
136. Kovari, A. Explainable AI Chatbots towards XAI ChatGPT: A Review. Heliyon 2025, 11, e42077. [CrossRef]
137. Hofeditz, L.; Clausen, S.; Rieß, A.; Mirbabaie, M.; Stieglitz, S. Applying XAI to an AI-Based System for Candidate Management to
Mitigate Bias and Discrimination in Hiring. Electron. Markets 2022, 32, 2207–2233. [CrossRef]
138. Veeramachaneni, V. Large Language Models: A Comprehensive Survey on Architectures, Applications, and Challenges. Adv.
Innov. Comput. Program. Lang. 2025, 7, 20–39. [CrossRef]
139. Bujold, A.; Roberge-Maltais, I.; Parent-Rocheleau, X.; Boasen, J.; Sénécal, S.; Léger, P.-M. Responsible Artificial Intelligence in
Human Resources Management: A Review of the Empirical Literature. AI Ethics 2024, 4, 1185–1200. [CrossRef]
140. Peng, K.; Peng, Y.; Li, W. Research on Customer Churn Prediction and Model Interpretability Analysis. PLoS ONE 2023,
18, e0289724. [CrossRef] [PubMed]
141. Guerranti, F.; Dimitri, G.M. A Comparison of Machine Learning Approaches for Predicting Employee Attrition. Appl. Sci. 2023,
13, 267. [CrossRef]
142. Gadgil, K.; Gill, S.S.; Abdelmoniem, A.M. A Meta-Learning Based Stacked Regression Approach for Customer Lifetime Value
Prediction. J. Economy Technol. 2023, 1, 197–207. [CrossRef]
143. Nawaz, N.; Arunachalam, H.; Pathi, B.K.; Gajenderan, V. The Adoption of Artificial Intelligence in Human Resources Management
Practices. Int. J. Inf. Manag. Data Insights 2024, 4, 100208. [CrossRef]
Algorithms 2025, 18, 556 34 of 36

144. Inavolu, S.M. Exploring AI-Driven Customer Service: Evolution, Architectures, Opportunities, Challenges and Future Directions.
Int. J. Eng. Adv. Technol. 2024, 13, 156–163. [CrossRef]
145. Polignano, M.; Musto, C.; Pellungrini, R.; Purificato, E.; Semeraro, G.; Setzu, M. XAI.it 2024: An Overview on the Future of AI in
the Era of Large Language Models. XAI.it 2024, 3839, 1–10.
146. Fantozzi, P.; Naldi, M. The Explainability of Transformers: Current Status and Directions. Computers 2024, 13, 92. [CrossRef]
147. Zhao, H.; Chen, H.; Yang, F.; Liu, N.; Deng, H.; Cai, H.; Wang, S.; Yin, D.; Du, M. Explainability for Large Language Models: A
Survey. ACM Trans. Intell. Syst. Technol. 2024, 15, 1–38. [CrossRef]
148. Wei, J.; Wang, X.; Schuurmans, D.; Bosma, M.; Ichter, B.; Xia, F.; Chi, E.; Le, Q.V.; Zhou, D. Chain of Thought Prompting Elicits
Reasoning in Large Language Models. arXiv 2022, arXiv:2201.11903. [CrossRef]
149. Lampinen, A.K.; Dasgupta, I.; Micheli, G.; Chan, S.; Bowman, H.; Hernandez, M.; Liu, J.Z.; Nematzadeh, A.; Schwettmann, S.;
Blundell, C.; et al. Can Language Models Learn from Explanations in Context? arXiv 2022, arXiv:2204.02329. [CrossRef]
150. Wiegreffe, S.; Pinter, Y. Attention Is Not Not Explanation. In Proceedings of the 2019 Conference on Empirical Methods in Natural
Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), Hong Kong,
China, 3–7 November 2019; Association for Computational Linguistics: Stroudsburg, PA, USA, 2019; Volume D19-1002, pp. 11–20.
[CrossRef]
151. Geva, M.; Schuster, T.; Berant, J.; Levy, O. Transformer Feed-Forward Layers Are Key-Value Memories. arXiv 2021,
arXiv:2106.05394. [CrossRef]
152. Jacovi, A.; Goldberg, Y. Towards Faithfully Interpretable NLP Systems: How Should We Define and Evaluate Faithfulness? In
Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL 2020), Online, 5–10 July 2020;
Association for Computational Linguistics: Stroudsburg, PA, USA; 2020; pp. 4198–4205. [CrossRef]
153. Zhang, Y.; Wang, X.; Wu, L.; Wang, J. Enhancing Chain of Thought Prompting in Large Language Models via Reasoning Patterns.
Proc. AAAI Conf. Artif. Intell. 2025, 39, 25985–25993. [CrossRef]
154. Saleh, Y.; Abu Talib, M.; Nasir, Q.; Dakalbab, F. Evaluating Large Language Models: A Systematic Review of Efficiency,
Applications, and Future Directions. Front. Comput. Sci. 2025, 7, 1523699. [CrossRef]
155. Patil, R.; Gudivada, V. A Review of Current Trends, Techniques, and Challenges in Large Language Models (LLMs). Appl. Sci.
2024, 14, 2074. [CrossRef]
156. Krause, S.; Stolzenburg, F. Commonsense Reasoning and Explainable Artificial Intelligence Using Large Language Models. In
Proceedings of the European Conference on Artificial Intelligence; Springer Nature Switzerland: Cham, Switzerland, 2023; pp. 302–319.
[CrossRef]
157. Randl, K.; Pavlopoulos, J.; Henriksson, A.; Lindgren, T. Evaluating the Reliability of Self-Explanations in Large Language Models.
In Proceedings of the International Conference on Discovery Science; Springer Nature Switzerland: Cham, Switzerland, 2024; pp. 36–51.
[CrossRef]
158. Chen, Y.; Liu, H.; Wen, Z.; Lin, W. How Explainable Machine Learning Enhances Intelligence in Explaining Consumer Purchase
Behavior: A Random Forest Model with Anchoring Effects. Systems 2023, 11, 312. [CrossRef]
159. Jain, N.; Jana, P.K. Xrrf: An Explainable Reasonably Randomised Forest Algorithm for Classification and Regression Problems.
Inf. Sci. 2022, 613, 139–160. [CrossRef]
160. Sayeed, M.A.; Rahman, A.; Rahman, A.; Rois, R. On the Interpretability of the SVM Model for Predicting Infant Mortality in
Bangladesh. J. Health Popul. Nutr. 2024, 43, 170. [CrossRef] [PubMed]
161. Loor, M.; Tapia-Rosero, A.; De Tré, G. Contextual Boosting to Explainable SVM Classification. In Fuzzy Logic and Technology, and
Aggregation Operators. EUSFLAT AGOP 2023; Massanet, S., Montes, S., Ruiz-Aguilera, D., González-Hidalgo, M., Eds.; Lecture
Notes in Computer Science; Springer: Cham, Switzerland, 2023; Volume 14069. [CrossRef]
162. de Waal, A.; Joubert, J.W. Explainable Bayesian Networks Applied to Transport Vulnerability. Expert Syst. Appl. 2022, 209, 118348.
[CrossRef]
163. Butz, R.; Schulz, R.; Hommersom, A.; van Eekelen, M. Investigating the Understandability of XAI Methods for Enhanced User
Experience: When Bayesian Network Users Became Detectives. Artif. Intell. Med. 2022, 134, 102438. [CrossRef]
164. Saleem, R.; Yuan, B.; Kurugollu, F.; Anjum, A.; Liu, L. Explaining Deep Neural Networks: A Survey on the Global Interpretation
Methods. Neurocomputing 2022, 513, 165–180. [CrossRef]
165. Dai, E.; Zhao, T.; Zhu, H.; Xu, J.; Guo, Z.; Liu, H.; Tang, J.; Wang, S. A Comprehensive Survey on Trustworthy Graph Neural
Networks: Privacy, Robustness, Fairness, and Explainability. Mach. Intell. Res. 2024, 21, 1011–1061. [CrossRef]
166. Otto, A.R.; Devine, S.; Schulz, E.; Bornstein, A.M.; Louie, K. Context-Dependent Choice and Evaluation in Real-World Consumer
Behavior. Sci. Rep. 2022, 12, 17744. [CrossRef]
167. Miró-Nicolau, M.; Jaume-i-Capó, A.; Moyà-Alcover, G. A Comprehensive Study on Fidelity Metrics for XAI. Inf. Process. Manag.
2025, 62, 103900. [CrossRef]
Algorithms 2025, 18, 556 35 of 36

168. Chen, J.; Song, L.; Yoon, J. Explaining Explanations: Axiomatic Feature Interactions for Deep Networks. In Proceedings of
the 32nd Conference on Neural Information Processing Systems (NeurIPS 2018), Montreal, QC, Canada, 3–8 December 2018.
[CrossRef]
169. Coroama, L.; Groza, A. Evaluation Metrics in Explainable Artificial Intelligence (XAI). In Proceedings of the International
Conference on Advanced Research in Technologies, Information, Innovation and Sustainability, Santiago de Compostela, Spain,
12–15 September 2022; Springer Nature Switzerland: Cham, Switzerland, 2022; pp. 401–413. [CrossRef]
170. Pearl, J. Causality: Models, Reasoning, and Inference, 2nd ed.; Cambridge University Press: Cambridge, UK, 2009.
171. Karimi, A.H.; Schölkopf, B.; Valera, I. Algorithmic Recourse: From Counterfactual Explanations to Interventions. In Proceedings
of the 2021 ACM Conference on Fairness, Accountability, and Transparency (FAccT 2021), Virtual Event, 3–10 March 2021;
Association for Computing Machinery: New York, NY, USA; 2021; pp. 353–362. [CrossRef]
172. Perotti, A.; Borile, C.; Miola, A.; Nerini, F.P.; Baracco, P.; Panisson, A. Explainability, Quantified: Benchmarking XAI Techniques.
In Proceedings of the World Conference on Explainable Artificial Intelligence, Valletta, Malta, 17–19 July 2024; Springer Nature
Switzerland: Cham, Switzerland, 2024; pp. 421–444. Available online: https://link.springer.com/chapter/10.1007/978-3-031-63
787-2_22 (accessed on 28 August 2025).
173. Bhattacharya, A. Applied Machine Learning Explainability Techniques: Make ML Models Explainable and Trustworthy for Practical
Applications Using LIME, SHAP, and More; Packt Publishing Ltd.: Birmingham, UK, 2022.
174. Merrick, L.; Taly, A. The Explanation Game: Explaining Machine Learning Models Using Shapley Values. In Machine Learning
and Knowledge Extraction—Proceedings of the 4th IFIP TC 5, TC 12, WG 8.4, WG 8.9, WG 12.9 International Cross-Domain Conference,
CD-MAKE 2020, Dublin, Ireland, 25–28 August 2020; Lecture Notes in Computer Science; Springer International Publishing: Cham,
Switzerland, 2020; Volume 12279, pp. 17–38. [CrossRef]
175. Molnar, C. Interpretable Machine Learning, 2nd ed.; Springer: Berlin/Heidelberg, Germany, 2020.
176. Drosou, M.; Jagadish, H.V.; Pitoura, E.; Stoyanovich, J. Diversity in Big Data: A Review. Big Data 2017, 5, 73–84. [CrossRef]
177. Agrawal, R.; Gollapudi, S.; Halverson, A.; Ieong, S. Diversifying Search Results. In Proceedings of the 2nd ACM International
Conference on Web Search and Data Mining (WSDM 2009), Barcelona, Spain, 9–11 February 2009; Association for Computing
Machinery: New York, NY, USA; 2009; pp. 5–14. [CrossRef]
178. O’Neil, C. Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy; Crown Publishing Group: New
York, NY, USA, 2016.
179. Chamola, V.; Hassija, V.; Sulthana, A.R.; Ghosh, D.; Dhingra, D.; Sikdar, B. A Review of Trustworthy and Explainable Artificial
Intelligence (XAI). IEEE Access 2023, 11, 78994–79015. [CrossRef]
180. High Level Expert Group on Artificial Intelligence, European Commission. Ethics Guidelines for Trustworthy AI; Technical Report;
European Commission: Brussels, Belgium, 2019. Available online: https://ec.europa.eu/newsroom/dae/document.cfm?doc_
id=60419 (accessed on 25 August 2025). [CrossRef]
181. Oh, S.J.; Benenson, R.; Fritz, M.; Schiele, B. Faceless Person Recognition: Privacy Implications in Social Media. In Proceedings of
the European Conference on Computer Vision (ECCV 2016), Amsterdam, The Netherlands, 8–16 October 2016; Springer: Cham,
Switzerland; 2016; pp. 19–35. [CrossRef]
182. Orekondy, T.; Schiele, B.; Fritz, M. Knockoff Nets: Stealing Functionality of Black-Box Models. In Proceedings of the IEEE/CVF
Conference on Computer Vision and Pattern Recognition (CVPR 2019), Long Beach, CA, USA, 15–20 June 2019; IEEE: Piscataway,
NJ, USA; 2019; pp. 4954–4963. [CrossRef]
183. Oh, S.J.; Schiele, B.; Fritz, M. Towards Reverse-Engineering Black-Box Neural Networks. In Explainable AI: Interpreting, Explaining
and Visualizing Deep Learning; Samek, W., Montavon, G., Vedaldi, A., Hansen, L.K., Müller, K.-R., Eds.; Springer: Cham,
Switzerland, 2019; pp. 121–144. [CrossRef]
184. Aditya, P.; Sen, R.; Druschel, P.; Oh, S.J.; Benenson, R.; Fritz, M.; Wu, T.T. I-PIC: A Platform for Privacy-Compliant Image Capture.
In Proceedings of the 14th Annual International Conference on Mobile Systems, Applications, and Services (MobiSys 2016),
Singapore, 25–30 June 2016; Association for Computing Machinery: New York, NY, USA; 2016; pp. 235–248. [CrossRef]
185. Sun, Q.; Tewari, A.; Xu, W.; Fritz, M.; Theobalt, C.; Schiele, B. A Hybrid Model for Identity Obfuscation by Face Replacement. In
Proceedings of the European Conference on Computer Vision (ECCV 2018), Munich, Germany, 8–14 September 2018; Springer:
Cham, Switzerland; 2018; pp. 553–569. [CrossRef]
186. Ziethmann, P.; Stieler, F.; Pfrommer, R.; Schlögl-Flierl, K.; Bauer, B. Towards a Framework for Interdisciplinary Studies in
Explainable Artificial Intelligence. In Proceedings of the International Conference on Human-Computer Interaction, Washington,
DC, USA, 29 June–4 July 2024; Springer Nature Switzerland: Cham, Switzerland, 2024; pp. 316–333. [CrossRef]
187. Vilone, G.; Longo, L. Notions of Explainability and Evaluation Approaches for Explainable Artificial Intelligence. Inf. Fusion 2021,
76, 89–106. [CrossRef]
188. Graziani, M.; Dutkiewicz, L.; Calvaresi, D.; Pereira Amorim, J.; Yordanova, K.; Vered, M.; Nair, R.; Henriques Abreu, P.; Blanke, T.;
Pulignano, V.; et al. A Global Taxonomy of Interpretable AI: Unifying the Terminology for the Technical and Social Sciences. Artif.
Intell. Rev. 2023, 56, 3473–3504. [CrossRef]
Algorithms 2025, 18, 556 36 of 36

189. Saeed, W.; Omlin, C. Explainable AI (XAI): A Systematic Meta-Survey of Current Challenges and Future Opportunities. Knowl.-
Based Syst. 2023, 263, 110273. [CrossRef]
190. Raees, M.; Meijerink, I.; Lykourentzou, I.; Khan, V.J.; Papangelis, K. From Explainable to Interactive AI: A Literature Review on
Current Trends in Human-AI Interaction. Int. J. Hum.-Comput. Stud. 2024, 180, 103301. [CrossRef]
191. Picard, R.W. Perceptual User Interfaces: Affective Perception. Commun. ACM 2000, 43, 50–51. [CrossRef]
192. Venkatasubbu, S.; Krishnamoorthy, G. Ethical Considerations in AI: Addressing Bias and Fairness in Machine Learning Models. J.
Knowl. Learn. Sci. Technol. 2022, 1, 130–138. [CrossRef]
193. Islam, M.U.; Mottalib, M.M.; Hassan, M.; Alam, Z.I.; Zobaed, S.M.; Rabby, M.F. The Past, Present, and Prospective Future of XAI:
A Comprehensive Review. In Explainable Artificial Intelligence for Cyber Security: Next Generation Artificial Intelligence; Springer:
Cham, Switzerland, 2022; pp. 1–29. [CrossRef]
194. Banabilah, S.; Aloqaily, M.; Alsayed, E.; Malik, N.; Jararweh, Y. Federated Learning Review: Fundamentals, Enabling Technologies,
and Future Applications. Inf. Process. Manag. 2022, 59, 103061. [CrossRef]
195. Hougen, D.F.; Shah, S.N.H. The Evolution of Reinforcement Learning. In Proceedings of the 2019 IEEE Symposium Series
on Computational Intelligence (SSCI), Xiamen, China, 6–9 December 2019; IEEE: Piscataway, NJ, USA, 2019; pp. 1457–1464.
[CrossRef]
196. Shokri, R.; Stronati, M.; Song, C.; Shmatikov, V. Privacy-Preserving Deep Learning. In Proceedings of the 2015 ACM SIGSAC
Conference on Computer and Communications Security (CCS 2015), Denver, CO, USA, 12–16 October 2015; Association for
Computing Machinery: New York, NY, USA; 2015; pp. 1310–1321. [CrossRef]
197. López-Blanco, R.; Alonso, R.S.; González-Arrieta, A.; Chamoso, P.; Prieto, J. Federated Learning of Explainable Artificial
Intelligence (FED-XAI): A Review. In Proceedings of the International Symposium on Distributed Computing and Artificial
Intelligence, Salamanca, Spain, 5–7 July 2023; Springer Nature Switzerland: Cham, Switzerland, 2023; pp. 318–326. [CrossRef]
198. Miró-Nicolau, M.; Jaume-i-Capó, A.; Moyà-Alcover, G. Assessing Fidelity in XAI Post-Hoc Techniques: A Comparative Study
with Ground Truth Explanations Datasets. Artif. Intell. 2024, 335, 104179. [CrossRef]
199. Lei, Y.; Li, Z.; Li, Y.; Zhang, J.; Shan, H. LICO: Explainable Models with Language-Image Consistency. In Proceedings of the
Thirty-Seventh Conference on Neural Information Processing Systems (NeurIPS 2023), New Orleans, LA, USA, 10–16 December
2023. Available online: https://arxiv.org/abs/2310.09821 (accessed on 28 August 2025).
200. Mhasawade, V.; Rahman, S.; Haskell-Craig, Z.; Chunara, R. Understanding Disparities in Post Hoc Machine Learning Explanation.
In Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency, Rio de Janeiro, Brazil, 3–6 June 2024;
pp. 2374–2388. [CrossRef]
201. Guidotti, R. Counterfactual Explanations and How to Find Them: Literature Review and Benchmarking. Data Min. Knowl. Discov.
2024, 38, 2770–2824. [CrossRef]
202. Pawelczyk, M.; Bielawski, S.; Heuvel, J.v.d.; Richter, T.; Kasneci, G. CARLA: A Python Library to Benchmark Algorithmic
Recourse and Counterfactual Explanation Algorithms. arXiv 2021, arXiv:2108.00783. [CrossRef]
203. Serrurier, M.; Mamalet, F.; Fel, T.; Béthune, L.; Boissin, T. On the Explainable Properties of 1-Lipschitz Neural Networks: An
Optimal Transport Perspective. Adv. Neural Inf. Process. Syst. 2023, 36, 54645–54682. Available online: https://arxiv.org/abs/2206
.06854 (accessed on 28 August 2025).
204. DeYoung, J.; Jain, S.; Rajani, N.F.; Lehman, E.; Xiong, C.; Socher, R.; Wallace, B.C. ERASER: A Benchmark to Evaluate Rationalized
NLP Models. arXiv 2019, arXiv:1911.03429. [CrossRef]
205. Wollek, A.; Graf, R.; Čečatka, S.; Fink, N.; Willem, T.; Sabel, B.O.; Lasser, T. Attention-Based Saliency Maps Improve Interpretability
of Pneumothorax Classification. Radiol. Artif. Intell. 2023, 5, e220187. [CrossRef]
206. Sadeghi, Z.; Alizadehsani, R.; Cifci, M.A.; Kausar, S.; Rehman, R.; Mahanta, P.; Bora, P.K.; Almasri, A.; Alkhawaldeh, R.; Hussain,
S.; et al. A Review of Explainable Artificial Intelligence in Healthcare. Comput. Electr. Eng. 2024, 118, 109370. [CrossRef]
207. Burger, C.; Walter, C.; Le, T. The Effect of Similarity Measures on Accurate Stability Estimates for Local Surrogate Models in
Text-Based Explainable AI. arXiv 2024, arXiv:2406.15839. [CrossRef]
208. Igami, M. Artificial Intelligence as Structural Estimation: Economic Interpretations of Deep Blue, Bonanza, and AlphaGo. arXiv
2017, arXiv:1710.10967. Available online: https://arxiv.org/abs/1710.10967 (accessed on 13 June 2025). [CrossRef]
209. United Nations General Assembly. Transforming Our World: The 2030 Agenda for Sustainable Development; United Na-
tions: New York, NY, USA, 2015. Available online: https://www.un.org/en/development/desa/population/migration/
generalassembly/docs/globalcompact/A_RES_70_1_E.pdf (accessed on 25 August 2025).

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.