Explainable Artificial Intelligence (XAI): Concepts, Taxonomies, Opportunities and Challenges toward Responsible AI

Motivation

• As opaque AI systems increasingly affect human lives, explainability is needed to make decisions understandable, justifiable, legitimate, and trustworthy in critical domains.DNNs are complex black-box models because their architectures contain hundreds of layers and millions of parameters.Source: §1 · p. 2
• Interpretability can help detect training bias, expose adversarial perturbations, and ensure that meaningful variables and truthful causality drive model outputs.Practical interpretation may explain mechanisms and predictions, visualize discrimination rules, or indicate what could perturb a model.Source: §1 · p. 2

XAI framework

• XAI seeks explainable models that retain learning performance while enabling humans to understand, appropriately trust, and effectively manage AI systems.Its approach also draws on Social Sciences and the psychology of explanation.Source: §1 · p. 2

Responsible AI

• The prospects extend XAI toward Responsible AI, requiring explainability together with fairness, accountability, privacy, robustness, and attention to confidentiality, security, data diversity, quality, integrity, and governance.The overview specifically calls for assessing XAI’s benefits and risks when sensitive information or confidential ML models are involved, including under data fusion paradigms.Source: §1 · pp. 3–4

Conceptual framework

• The overview proposes a novel explainability definition centered on the audience and purposes such as trustworthiness and privacy awareness, alongside transparency levels and post-hoc approaches.It distinguishes models transparent by design from explanations of models that are not transparent by design.Source: §1 · p. 3

Literature taxonomies

• It analyzes approximately 400 XAI contributions through taxonomies covering transparent and post-hoc ML models, plus Deep Learning methods organized by criteria such as layers, representation vectors, and attention.The overview also identifies unresolved needs for explainability concepts and metrics and for making Deep Learning models more understandable.Source: §1 · pp. 3–4

2.1. Terminology Clarification

• Explainability differs from interpretability: interpretability is a model’s passive human understandability, whereas explainability is an active procedure clarifying internal functions.Transparency denotes models understandable by themselves, while explainability can include post-hoc techniques that make non-interpretable models explainable.Source: §2.1 · pp. 4–5, §2.2 · p. 6

2.2. What?

• Given an audience, explainability is the details and reasons a model produces to make its functioning clear or easy to understand.The definition treats audience cognitive skills and goals as central, while recognizing that purposes include trustworthiness and other application needs.Source: §2.2 · pp. 6–7
• A general metric for XAI quality remains an open challenge because complexity reductions are quantifiable, whereas visualization and natural-language gains are harder to measure.The section notes that explanation quality depends on whether the intended audience has understood the model’s functioning.Source: §2.2 · p. 6

2.3. Why?

• Explainability is motivated by deployment barriers in regulated sectors and by the need for understanding beyond performance, supporting model trust, knowledge, and practical utility.Different audiences seek understanding, scientific knowledge, compliance, audits, or verification of decisions.Source: §2.3 · p. 7, Figure 2 · p. 7

2.4. What for?

• The reviewed literature identifies trustworthiness, causality, transferability, informativeness, confidence, fairness, accessibility, interactivity, and privacy awareness as explainability goals.Informativeness is the most frequently stated rationale, transferability the second, and accessibility the third among the surveyed contributions; causality alone cannot establish cause-effect relationships.Source: §2.4 · pp. 7–9

2.5. How?

• XAI approaches divide into transparent models interpretable by design and post-hoc techniques that explain otherwise opaque models.Post-hoc categories include text explanations, visualizations, local explanations, examples, simplification, and feature relevance.Source: §2.5 · p. 10

2.5.1. Levels of Transparency in Machine Learning Models

• Transparency comprises algorithmic transparency, decomposability, and simulatability, with each higher level containing its predecessors.Simulatability requires a human to reason about the model as a whole; decomposability requires understandable inputs, parameters, and calculations; algorithmic transparency requires mathematical analysis of the process.Source: §2.5.1 · pp. 10–11, Figure 3 · p. 11

2.5.2. Post-hoc Explainability Techniques for Machine Learning Models

• Post-hoc explainability targets models that are not interpretable by design, classifying techniques by explanatory intention, method, and applied data type for later ML-model reviews.The classification is presented in Figure 4 and used to organize the techniques reviewed in subsequent sections and Table 2.Source: §2.5.2 · pp. 11–12
• Text explanations generate natural-language or symbolic representations that communicate a model’s results or rationale through semantic mappings.These methods learn to generate text explanations and may use symbols to portray algorithmic functioning.Source: §2.5.2 · p. 12
• Visual explanations depict model behavior, often using dimensionality reduction, and can expose complex variable interactions to users unfamiliar with ML modeling.Visualizations may also be coupled with other techniques to improve understanding.Source: §2.5.2 · p. 12
• Local explanations explain only relevant, less-complex subspaces of a model, whereas example-based explanations extract representative data examples that clarify learned relationships and correlations.Both approaches focus on part of the system or on examples related to a generated result rather than explaining the entire model directly.Source: §2.5.2 · p. 12
• Simplification methods rebuild a lower-complexity system that resembles the original model while maintaining similar performance, while feature-relevance methods quantify each variable’s influence through relevance scores.Simplified models are generally easier to implement, and comparing feature scores reveals the importance assigned to variables in producing outputs.Source: §2.5.2 · p. 12

3.1. Linear/Logistic Regression

• Linear and logistic regression are transparent through stiffness, decomposability, simulatability, and algorithmic transparency, yet interpretation depends on audience and can be misleading across models, samples, groups, or time.Interpretations of odds ratios may conflate substantive effects with unobserved heterogeneity, while model size and engineered features can undermine simulatability and decomposability.Source: §3.1 · pp. 13–14

3.2. Decision Trees

• Decision trees can satisfy all transparency levels, but larger trees, complex feature relations, and tree ensembles progressively reduce human interpretability and may require post-hoc explanations.Small trees are simulatable, larger trees become decomposable, and sufficiently complex trees are only algorithmically transparent; ensembles lose transparent properties while improving generalization.Source: §3.2 · p. 15

3.3. K-Nearest Neighbors

• K-Nearest Neighbors explain predictions through similar past cases, allowing users to inspect neighborhood reasons and changes as K varies, although complex features, distances, or large K reduce transparency.Its transparency level depends on the features, number of neighbors, and distance function used.Source: §3.3 · p. 16

3.4. Rule-based Learning

• Rule-based learners provide interpretable linguistic or conditional representations, but increasing rule coverage and specificity improves performance while moving models toward algorithmic transparency.Fuzzy rules can handle imprecision and relax rule-size constraints; sufficiently broad rule wrappers can explain model behavior to non-experts while remaining standalone predictors.Source: §3.4 · p. 16

3.5. General Additive Models

• Generalized Additive Models expose each variable’s contribution through smooth functions and are adopted where understanding relationships matters more than maximizing accuracy.Their transparency can support practical use across finance, environmental studies, geology, healthcare, biology, and energy, although link functions or interactions may limit simulatability and decomposability.Source: §3.5 · p. 17

3.6. Bayesian Models

• Bayesian models transparently represent feature–target relationships through probabilistic graphical connections, but overly complex variables can remove simulatability and decomposability.They support insights across domains and have also been used to explain other models, including averaged tree ensembles.Source: §3.6 · p. 17

4.1 Model-agnostic techniques

• Model-agnostic post-hoc techniques explain predictions without depending on a model’s internal processing or representations.These methods can be applied seamlessly across different machine-learning models.Source: §4 · p. 18

4.2 Shallow models

• Post-hoc methods for shallow models address learners without layered neural-processing structures and identify trends across recent contributions.The review covers techniques specifically designed for shallow machine-learning models.Source: §4 · p. 18

4.3 Deep learning

• Deep-learning explainability methods target neural networks and variants, including convolutional, recurrent, and hybrid models combining deep networks with transparent models.The section reviews recent methods and trends for each model family.Source: §4 · p. 18

4.4 Deep-learning taxonomy

• A complementary taxonomy classifies post-hoc explanations for deep-learning models by linking their specific aspects to criteria from the general taxonomy.This analysis focuses on characteristics of deep-learning black boxes.Source: §4 · p. 18

Taxonomy

• Model-agnostic post-hoc methods extract information from any model through simplification, feature relevance estimation, or visualization.Some methods build reduced-complexity proxies, others extract knowledge directly, and others visualize model behavior.Source: §4.1 · p. 18

Explanation by simplification

• Rule extraction dominates model-agnostic contributions, with LIME producing locally linear models around opaque-model predictions and related simplification methods building interpretable proxies.Other approaches include G-REX, CNF/DNF rule learning [215], model extraction, and black-box distillation and auditing.Source: §4.1 · pp. 18–20

Feature relevance

• Feature relevance methods quantify how individual features influence predictions through additive scores, game theory, gradients, sensitivity analysis, influence functions, and counterfactual inputs.SHAP provides additive feature-importance scores with local accuracy, missingness, and consistency, while QII accounts for correlated inputs.Source: §4.1 · pp. 19–20

Visualization

• Visual explanations are less common because model-agnostic visualizations must rely only on opaque-model inputs and outputs, so they typically display feature-relevance information.Examples include global sensitivity-analysis visualizations and ICE plots.Source: §4.1 · pp. 20–21

4.2.1. Tree Ensembles, Random Forests and Multiple Classifier Systems

• Tree-ensemble explanations mainly use simplification and feature relevance, with limited work on boosting and stacking compared with bagging ensembles.Simplification trains interpretable surrogates or pairs simple and complex models [118] [122], while feature relevance includes Random Forest variable importance and recommendation-oriented counterfactuals. Stacking With Auxiliary Features links component predictions to their ensemble contribution.Source: §4.2.1 · pp. 21–22

4.2.2. Support Vector Machines

• SVM post-hoc explainability spans rule-based simplification, local explanations, visualizations, and examples because SVMs have comparatively opaque mathematical structures.Rule extraction progresses from support-vector rules to hyperplane-based, prototype-based, multi-constrained, and Voronoi-based representations. Visual methods expose kernel information, input-variable relevance, compound modifications, and margin-based multivariate patterns.Source: §4.2.2 · pp. 22–23
• In SVMs, simplification methods are generally older than emerging local explanations, and the reviewed explainability literature ends in 2017.Bayesian interpretations make hyperparameter tuning and class probabilities more comprehensible, while geometrical analysis of indefinite kernels challenges the usual hyperplane-margin interpretation.Source: §4.2.2 · p. 23

4.3. Explainability in Deep Learning

• Deep-learning explainability primarily relies on post-hoc local explanations and feature-relevance techniques across multi-layer networks, CNNs, and RNNs.These approaches address the practical reluctance caused by neural networks’ black-box character.Source: §4.3 · p. 23, §4.3.1 · p. 23

4.3.1. Multi-layer Neural Networks

• Feature-relevance methods for multi-layer networks include deep Taylor decomposition and DeepLIFT, while integrated gradients satisfies sensitivity and implementation invariance axioms.PatternNet and PatternAttribution were introduced after analyses found that DeConvNet, Guided BackProp, and LRP could produce theoretically incorrect explanations.Source: §4.3.1 · p. 24

4.3.2. Convolutional Neural Networks

• CNN explanations map predictions to discriminative image regions or interpret internal representations using saliency maps, activation visualizations, attribution heatmaps, and class activation methods such as Grad-CAM.CNN explainability also includes text descriptions, network dissection, feature visualization, and simpler local surrogate explanations such as LIME.Source: §4.3.2 · pp. 24–27

4.3.3. Recurrent Neural Networks

• RNN explainability remains comparatively limited and combines feature-relevance analyses with architecture modifications that expose influential patterns or interpretable latent representations.Methods include LRP adaptations for LSTMs and GRUs, finite-horizon n-gram visualizations, interpretable mimic learning, RETAIN attention, SISTA-based RNNs, and HMM-RNN combinations.Source: §4.3.3 · p. 27

4.3.4. Hybrid Transparent and Black-box Methods

• Hybrid transparent and black-box methods inject symbolic or probabilistic knowledge, use interpretable neighbors or twins, and can improve explainability, performance, and robustness to mislabeled training data.Examples include knowledge-base constraints, probabilistic graphical-model hybrids, DkNN, white-box ensembles, and CBR systems paired with DNNs while preserving accuracy.Source: §4.3.4 · pp. 27–28

Alternative Deep Learning taxonomy

• The alternative taxonomy groups methods into explanations of Deep Network Processing, explanations of Deep Network Representation, and explanations by simplification [13].Processing asks why an input produces an output; representation asks what information the network contains; simplification targets models designed to make their behavior interpretable.Source: §4.4 · pp. 28–29

Taxonomy overlap

• Figure 11 shows that methods separated by feature type in Figure 6 can form one Salience Mapping category for explaining Deep Network Processing [272] [280].This demonstrates that taxonomy placement depends on the classification perspective.Source: Figure 11 · p. 29, §4.4 · p. 29
• Figure 11 also splits methods grouped under simplification in Figure 6 between Decision Trees for network processing and Representation Vectors for network representation [144].Thus, a single method category in one taxonomy may correspond to multiple categories in another.Source: Figure 11 · p. 29, §4.4 · p. 29

Method selection

• Separating model processing from model representation distinguishes execution traces from internal data structures and supports choosing an XAI method based on the failure information required.The distinction parallels testing and debugging in conventional programming.Source: §4.4 · p. 29

5. XAI: Opportunities, Challenges and Future Research Needs

• The prospective agenda revisits XAI challenges and research opportunities, complemented by broader responsible-AI research needs and open questions.The discussion also points toward underexplored niches connected to model explainability.Source: §5 · pp. 29–30

5.1. On the Tradeoff between Interpretability and Performance

• The interpretability–performance tradeoff is conditional: greater complexity may help when functions and data are sufficiently complex and diverse, but not otherwise.Complex models provide greater flexibility, whereas controlled, high-quality features can make simpler models equally or more accurate [347].Source: §5.1 · p. 30
• XAI methods may improve the usual interpretability–performance tradeoff rather than leaving interpretability on an unavoidable downward slope.Figure 12 presents this potential area of improvement, inspired by previous work.Source: §5.1 · p. 30, Figure 12 · p. 31
• Explanations must balance approximation and fidelity, remaining understandable to their audience without oversimplifying essential model features.This is presented as an approximation dilemma linked to model interpretability and performance.Source: §5.1 · p. 30

5.2. On the Concept and Metrics

• XAI needs a unified concept defining explainability as a model’s ability to make its functioning clearer to an audience, providing common ground for the field.The proposed concept is intended to support a common structure for XAI systems, although the authors acknowledge it may be incomplete.Source: §5.2 · p. 31
• Explainability requires metrics that compare how well models satisfy the concept by assessing explanation quality, usefulness, audience understanding, fidelity, trust, reliance, and model impact.Existing approaches include goodness checklists, satisfaction scales, mental-model elicitation, explainer-fidelity measures, explanation trustworthiness, and model reliability [349] [350].Source: §5.2 · p. 31
• More quantifiable and general XAI metrics, plus comparison methodologies across contexts, models, and purposes, remain necessary for evaluating XAI techniques.The survey does not design this metric suite, arguing that the broader community should address it after establishing a shared explainability concept.Source: §5.2 · pp. 31–32

5.3. Challenges to achieve Explainable Deep Learning

• 5.3: XAI for deep learning lacks standardized terminology and objective evaluation criteria, while explanation quality depends on audience and should emphasize causal, selective, and counterfactual reasoning [6] [13].The paper suggests evaluating explanations consistently with social-science methods and tailoring intelligibility to experts, policy-makers, and nontechnical users.Source: §5.3 · p. 32
• 5.3: Neural-symbolic interpretability can combine the precision of connectionist models with the greater explainability of symbolic reasoning, potentially preserving or improving generic performance [299] [312] [353].Symbolic knowledge bases support qualitative processing and established reasoning rules, while semantic representations can produce human-understandable or natural-language explanations.Source: §5.3 · p. 32
• 5.3: Explainable deep learning must provide accessible explanations for society, policy-makers, and law, including requirements related to ambiguities and the GDPR right to explanation.Nontechnical explanations are presented as necessary for broader social and legal acceptance.Source: §5.3 · p. 33

5.4. Explanations for AI Security: XAI and Adversarial Machine Learning

• 5.4: XAI can expose confidential model knowledge and enable model-functionality stealing, creating a need for explanation methods that preserve intellectual property and trade secrets [356].Even input-output access may compromise knowledge synthesized through years of research.Source: §5.4 · p. 33
• 5.4: XAI disclosures can strengthen adversarial attacks, such as minimal input changes that alter classification, while tools including CleverHans, AlfaSVMLib, and AdversarialLib target model hardening and vulnerability detection [358–363].The paper cites an autonomous-vehicle example in which an imperceptible stop-sign change caused detection as a 45 mph signal.Source: §5.4 · p. 33
• 5.4: Generative models can support explainability by linking latent representations to predicted classes and generating counterfactual prototypes that reveal model boundaries and enable informed criticism [368].GANs, variational autoencoders, and related models are presented as attribution and counterfactual-generation methods.Source: §5.4 · p. 33

5.5. XAI and Output Confidence

• 5.5: Because erroneous outputs in safety-critical domains can cause harm, output-confidence analysis and XAI-based inspection of attended input regions can expose epistemic uncertainty and support rejection of unreliable predictions.Relevant domains include autonomous vehicles, surgery, medical diagnosis, insurance risk assessment, and manufacturing cyber-physical systems.Source: §5.5 · pp. 33–34

5.6. XAI, Rationale Explanation, and Critical Data Studies

• XAI supports collaborative sense-making by communicating data practices, model reasoning, and findings across Data Science and non-expert disciplinary audiences.This includes sharing narratives about how data were collected, cleaned, modeled, and analyzed.Source: §5.6 · p. 34
• Rationale explanation makes model decisions accessible to non-experts and strengthens knowledge exchange in multidisciplinary Critical Data Studies.The approach helps teams appraise the ethical implications of data-based choices.Source: §5.6 · p. 34

5.7. XAI and Theory-guided Data Science

• Theory-guided Data Science aligns model selection, structure, training, regularization, and outputs with domain knowledge and theoretically plausible relations.Its goal is for models to generate knowledge that can be merged with what the field already knows.Source: §5.7 · p. 34
• Theory-guided Data Science and XAI converge because models should be both explainable and knowledge-consistent, with learned knowledge assessed against prior theoretical principles.Applications across diverse fields demonstrate the paradigm’s promising potential -.Source: §5.7 · pp. 34–35

5.8. Guidelines for ensuring Interpretable AI Models

• Interpretable AI requires considering the interests and requirements of all stakeholders, from system designers to decision makers and affected users.The process is multidisciplinary and inclusive rather than limited to model developers.Source: §5.8 · p. 35
• The guidelines in recommend context-sensitive design, preference for interpretable techniques when possible, and risk-aware supplementation with XAI when black-box models are necessary.Model choice should reflect domain risks, data resources, knowledge, task suitability, and whether opaque models provide superior capabilities for the problem.Source: §5.8 · p. 35
• Interpretability should match human cognitive capacities and support responsible AI through attention to mental models, accessibility, fairness, privacy, sustainability, and accountability [383, 384].For black-box systems, teams should evaluate explanatory coverage, define delivery strategies, and assign implementation responsibilities.Source: §5.8 · pp. 35–36

Toward Responsible AI

• Responsible AI implementation must address explainability alongside other principles, especially fairness and privacy.The section connects XAI and data fusion to the fulfillment of Responsible AI principles in practice.Source: §6 · p. 36

6.1. Principles of Artificial Intelligence

• AI-principles reviews classify organizations by nature, principle content and coverage, target audience, and relationship to International Human Rights.The content dimension includes principles such as privacy, explainability, and fairness, while the target audience may be internal or external.Source: §6.1 · p. 36
• The illustrative five-principle framework combines Fair AI, Transparent and Explainable AI, Human-centric AI, and Privacy and Security by Design.It links fairness to nondiscrimination, explainability to audience-adapted transparency, human-centricity to the Sustainable Development Goals and AI for Social Good [388], and privacy and security to Responsible Research and Innovation.Source: §6.1 · pp. 36–37
• The European Commission’s Trustworthy AI guidelines assess human agency, robustness, privacy, transparency, fairness, societal well-being, and accountability across AI organizations.Their checklist can be completed by product managers, developers, and other roles, extending the principles to organizations involved in AI development.Source: §6.1 · p. 37
• Most reviewed AI-principles guides explicitly include XAI: 28 of 32 identify it as a crucial component.This positions explainability as a major worldwide concern in AI-principles guidance.Source: §6.1 · p. 37

6.2.1. Fairness and Discrimination

• XAI supports fairness by revealing protected-feature correlations, discriminatory outcomes, and bias sources, while enabling counterfactual explanations and quantitative bias reporting.Bias may arise from skewed or tainted data, limited features, sample-size disparities, or proxy features; fairness can be evaluated and mitigated through pre-, in-, and post-processing techniques.Source: §6.2.1 · p. 38
• Fairness is assessed through independence, separation, and sufficiency criteria, although these criteria cannot all be fulfilled simultaneously and must instead be jointly optimized.Evaluation uses metrics such as statistical parity difference, disparate impact, equal opportunity difference, average odds difference, and the Theil index; mitigation includes pre-processing, in-processing, and post-processing methods.Source: §6.2.1 · pp. 38–39
• Fairness and XAI are commonly addressed together: 26 of 28 surveyed AI principles dealing with XAI also explicitly discuss fairness.Fair-by-design and counterfactual approaches can jointly reduce bias, provide human-comprehensible explanations, assess robustness, and examine individual and group fairness.Source: §6.2.1 · pp. 39–40

6.2.2. Accountability

• XAI contributes to accountability by making AI systems more auditable for different profiles, including regulators, and by supporting the minimization and reporting of negative impacts.Accountability also requires impact assessment, documented and ethically justified trade-offs, external auditing for systems affecting fundamental rights, and redress mechanisms for unforeseen unjust harms.Source: §6.2.2 · pp. 40–41

6.3. Privacy and Data Fusion

• Data fusion may improve ML performance across applications, but its intersection with XAI and privacy remains largely uncharted.The section frames data fusion as simultaneously offering explainability opportunities and creating privacy risks in Responsible AI.Source: §6.3 · p. 41

6.3.1. Basic Levels of Data Fusion

• Data fusion is categorized by data, model, or knowledge level, and by centralized processing or distributed local processing.Data-level fusion combines raw inputs, model-level fusion aggregates learned models, and knowledge-level fusion merges rules, ontologies, or other representations.Source: §6.3.1 · p. 41

6.3.2. Emerging Data Fusion Approaches

• Emerging approaches distribute or jointly optimize fusion: Big Data fusion maps local learning to workers, Federated Learning aggregates encrypted updates without sharing client data, and Multiview Learning jointly optimizes multiple views.Federated Learning also moves computation closer to collected data, reducing central-server burden and processing latency; Multiview Learning targets generalization under weak supervision.Source: §6.3.2 · p. 42

6.3.3. Opportunities and Challenges in Privacy and Data Fusion under the Responsible AI Paradigm

• Responsible AI requires privacy throughout the lifecycle, including data protection, quality and integrity, governance, and controlled access because privacy is intertwined with fairness.Collected and generated personal information can enable discrimination or unfair treatment when misused.Source: §6.3.3 · pp. 42–43
• XAI can explain how deep models fuse sources, improving model understanding and guiding additional sources or more efficient fusion, but explanations may reveal protected information.Feature-correlation analysis can expose task-induced spatial or temporal relationships, while inferred ownership or hidden features can compromise confidentiality.Source: §6.3.3 · pp. 43–44
• Distributed and knowledge-level fusion introduce further risks because shared model updates or extracted explanations may enable reconstruction of private context, differential-privacy breaches, or stronger adversarial attacks.Federated Learning avoids sharing raw data, yet post-hoc explanations and black-box queries can expose local training information or model hyperparameters.Source: §6.3.3 · p. 44

• Responsible AI principles divide into AI-specific concerns, such as explainability, fairness, and human agency, and end-to-end concerns including privacy, security, and safety.The EC Guidelines for Trustworthy AI exemplify end-to-end principles, whereas Telefonica’s principles are more AI-specific [386].Source: §6.4 · p. 45
• Implementation should operationalize AI-specific principles while leveraging existing organizational processes for generic principles such as privacy, security, and safety.A methodology such as [386] breaks implementation into coordinated organizational components.Source: §6.4 · p. 45
• A practical methodology combines principles, awareness and training, impact questionnaires, mitigation tools, and governance assigning responsibilities and accountabilities.Relevant tools include XAI and fairness tools, as well as model cards.Source: §6.4 · p. 45
• For agile organizations, employee awareness and self-responsibility should generally guide governance, with committees reserved for specific important issues.The alternative governance philosophies are committee review and approval versus employee self-responsibility.Source: §6.4 · p. 45
• Responsible AI implementation must balance major cultural and organizational change with feasibility and compliance across existing IT assets, policies, and resources.The gradual rise of corporate awareness is identified as a pathway for XAI to create substantial impact.Source: §6.4 · p. 45

Literature review

• The review distinguishes transparent ML models from post-hoc XAI techniques and develops taxonomies for both general ML proposals and Deep Learning domains.The Deep Learning taxonomy connects explainability methods more closely to the domains where they can be realized.Source: §7 · p. 46

Responsible AI

• XAI is presented as a core component of Responsible AI, whose practical principles include fairness, transparency, and privacy.Figure 14 summarizes the overview’s XAI challenges and their impact on Responsible AI principles.Source: §7 · p. 46, Figure 14 · p. 46

Challenges

• The overview highlights that XAI can compromise protected-data privacy during data fusion and has broader implications for fairness.These issues motivate treating explainability as part of a wider responsible-AI framework rather than as an isolated model property.Source: §7 · p. 46

Outlook

• Future XAI work must jointly consider model interpretability, data privacy, model confidentiality, fairness, and accountability.The authors argue that responsible AI implementation in organizations and institutions depends on studying these principles together.Source: §7 · p. 46