Abstract

Nowadays, artificial intelligence (AI) seems to have an answer for almost everything across major fields such as healthcare and retail. However, explainability is a significant weakness of AI. While modern AI delivers outstanding performance, its “black box” creates significant trust and safety difficulties for users. This paper provides an overview of explainable AI (XAI), a field dedicated to making AI decisions more transparent and interpretable. The paper outlines the historical transition to deep learning (DL) and defines the key concepts, advantages and limitations of XAI. It explores how XAI works, compares white box and black box models, and categorises explainability techniques by timing, scope and implementability. This paper also examines the applications of XAI in healthcare, retail and autonomous vehicles. Finally, the research paper discusses key ethical considerations, including fairness and data privacy, and outlines current implementation challenges and future trends in XAI.

1. Introduction

1.1 Introduction to Artificial Intelligence

Artificial intelligence (AI) can be defined as the science and engineering of creating intelligent machines and computer programs that simulate cognitive functions (McCarthy, 2007). Although AI has evolved rapidly in recent years, the pursuit of creating such systems began decades ago. One of the early landmark papers was by Alan Turing in 1950, which considered the question “can machines think?”. This paper gave us the “imitation game” (Turing, 1950) and motivated a large part of the research in the field. A more formal launch of what we came to call AI was at the 1956 Dartmouth Summer Research Project at Hanover (McCarthy et al., 1955). AI began with early approaches emphasising logical methods and rules; however, this simple rule-based system failed to account for complex problems.

The shortcomings of logic systems, such as they were unable to carry out even simple tasks which required knowledge about the physical world, prompted a movement towards machine learning (ML). Instead of coding the desired behaviour using pre-programmed rules, ML involves designing systems that can identify and use patterns in the training data, automatically adjusting them based on this information (Jordan & Mitchell, 2015). ML transformed AI from a logical rule-based approach into data-driven intelligent systems that improve performance on tasks through exposure to more data. The success of ML paved the way for deep learning (DL), which has proven remarkably effective by using a variety of learning models that utilise multi-layered neural network structures (LeCun et al., 2015) to find patterns and represent features from raw data, including images, sounds and text, making it important in applications of speech and image recognition. Overall, AI defines the objective of creating intelligent systems, ML enables these systems to recognise patterns from data and DL achieves this by using deep neural networks (DNNs).

However, recent studies on DL systems have drawn attention to their transparency, with networks being described as deep and opaque as a result of the complex multi-layered structure (LeCun et al., 2015). While early symbolic approaches were easily interpretable but struggled with complex real-world tasks, modern DL achieves superior performance by introducing black-box complexity. Consequently, understanding and mitigating this opacity has become more difficult, which emphasises the critical demand for explainability in modern AI systems.

1.2 Explainable Artificial Intelligence

Explainable artificial intelligence (XAI) refers to a set of methods and techniques that are designed to assist human users in understanding, interpreting and appropriately trusting the outputs of AI models (Arrieta et al., 2020). Instead of leaving the model’s reasoning hidden, XAI aims to show the reasoning behind a prediction in a way that people can understand, whether they are developers, domain experts or everyday users (Guidotti et al., 2018).

XAI is built as a field concerned with improving the transparency of AI systems overall. As predictive accuracy improved through DL and other advanced methods, models became harder to understand, making it difficult even for their own developers to explain how a specific input led to a specific output (Arrieta et al., 2020). This concern increased research investments, most notably DARPA’s XAI program, which was created specifically to build AI systems that can explain their decisions clearly while still performing well (Gunning & Aha, 2019).

To understand why XAI is necessary, it is useful to distinguish between black-box and explainable models – the major difference being whether the user can follow the reasoning of an output. Black-box models, such as deep neural networks and large ensemble methods, achieve strong performance but make the relationship between the inputs and outputs difficult to trace, giving no clear view into how a decision was made (Guidotti et al., 2018). Explainable AI, on the other hand, either uses models that are naturally easy to understand or use additional tools and techniques to reveal how a black-box model reached its decision (Mersha et al., 2024).

Explainability also affects trust, accountability and usability once a system is deployed. Different stakeholders rely on it for different reasons, such as users verifying that a decision was fair, organisations justifying decisions internally and to regulators, and regulators confirming compliance with legislation (Arrieta et al., 2020). According to NIST’s four principles of explainable AI, a system’s output should: provide its reasons or evidence; be meaningful to the person using it; accurately reflect how the model actually works; and recognise when it’s operating outside of its knowledge (Phillips et al., 2021). Without these qualities, people cannot check whether a decision was fair, correct or safe to act on. This is especially important in high-stakes areas such as healthcare, credit and law, where a wrong or unexplained decision can have serious consequences (Mersha et al., 2024).

1.3 Advantages and Key Features of XAI

The concept of explainability in AI proves to be necessary in countless areas as it addresses some of the principal concerns behind AI, which would indisputably be transparency, accountability and fairness. This is due to its operational mechanics and technical complexity being difficult for the majority of users, predominantly due to the black-box nature where transparency is lacking as a result of complex non-linear algorithms (Dwivedi et al., 2023). However, this lack of transparency and accountability could prove to be fatal when they are not upheld in fields like healthcare, military and finance. 

XAI plays an important role in bridging the gap between human trust and modern ML through providing comprehensible explanations for users (Knap, 2022). From this statement we can infer that XAI provides the justification for predicted output, enabling users to understand the prediction. This can be implemented in sales where users are able to understand price recommendations due to explanations provided by XAI. Additionally, workers also benefit from this technology by being able to monitor any bias, discrimination and errors to maintain accountability (Novelli, 2023). This transparency helps organisations examine causes of discrimination and any errors that could appear in a prediction, hence eliminating these blunders for future predictions (Arrieta, 2020). These elements have been in high demand since the adoption of AI in high-stakes domains because users need to understand and justify automated decisions. 

Aside from the advantages of XAI, there are some significant technical features of XAI like global and local explanations. A global explanation explains the feature importance across the entire system, while a local explanation describes the feature importance for a prediction on an individual instance (Salih et al., 2024). Global and local explanations are provided by SHapley Additive exPlanations (SHAP), initially invented in 2017 by Lundberg and Lee. It uses game theory to assign values to each feature and then weighs the importance of these features on the final prediction. Local Interpretable Model-agnostic Explanations (LIME), invented in 2016, seeks to provide individual explanations by fitting a local surrogate model to explain the behaviour of a complex model (Salih et al., 2024). Another important explanation technique is feature importance explanations which summarise feature significance in model analysis (Huang et al., 2022). These distinct techniques and advantages allow various practitioners to increase development in accurate, reliable and efficient systems that support human interpretability and trust (Kaya et al., 2026).

1.4 Challenges in XAI

Although XAI has improved the transparency of modern systems by addressing the limitations of the black-box model, several challenges reduce its practicality in real-world adoption (Kaya et al., 2026; Kalasampath et al., 2025). These challenges become more prominent in complex systems such as black-box models (Kaya et al., 2026).

One prominent challenge XAI may struggle with is balancing predictive accuracy and model interpretability. Interpretability refers to the extent to which humans can understand the reasoning behind an AI model’s predictions. As AI models become more complex, it becomes difficult to produce high-quality explanations. This challenge is illustrated in a 2026 survey, which shows that ante-hoc models such as decision trees integrate interpretability into their architecture, making it simpler for users to understand how predictions are deduced. However, these models underperform compared to black-box models and more complex programs than decision trees, such as DNNs. To overcome this, techniques such as SHAP and LIME (Kaya et al., 2026) are implemented to generate explanations for large models. As a result of its increased scalability, this will eventually lead to a rise in costs.

Moreover, XAI may struggle with increasing computational cost associated with maintaining scalability as AI models become more complex. With the need for more advanced systems, XAI methods must also become increasingly sophisticated to generate meaningful explanations for larger and more complex models. This results in higher computational costs, particularly when integrating explanation techniques such as SHAP and LIME into existing AI systems (Kaya et al., 2026). Furthermore, as the size and complexity of AI models increase, generating explanations becomes more computationally intensive, reducing the practicality of XAI in real-time and safety-critical applications where rapid decision-making is essential (Kaya et al., 2026). Although approximation techniques can reduce computational overhead, they may also reduce the accuracy and faithfulness of the explanations provided, creating a trade-off between efficiency and explanation quality (Kaya et al., 2026). While XAI improves transparency, its computational demands and scalability may limit its adoption in large-scale systems.

These findings show that maintaining interpretability may become increasingly difficult as models become more scalable in size which results in a decline in the accuracy of the model. This could raise  issues regarding its efficiency and its ability to interpret large databases to deliver a response that readers would be able to understand. It is essential that XAI should be able to maintain these aspects if it were to become increasingly scalable.

2. How Does XAI Work?

To understand how XAI functions, one should first consider how various types of models of AI decide. Models of AI tend to fall into two categories: white-box models (also called glass-box models) and black-box models (Rudin, 2019). White-box models (like linear regression, trees and simple rule-based models) are inherently transparent models that allow a specialist to look at their coefficients to recognise just how an input is transformed  into an output. Black-box designs (like neural networks and support vector machines), however, are significantly more complicated. 

These models can offer high degrees of accuracy in intricate jobs, but are opaque, providing no visibility  into their internal workings. In such scenarios, there is often an inherent trade-off between the accuracy of the model and its interpretability (Arrieta et al., 2020). The various methods utilised in XAI are usually classified into three types based on three criteria (Arrieta et al., 2020):

• First, the concept of time – intrinsic explainability refers to methods based on models that are by their nature explainable, whereas post-hoc explainability refers to techniques for creating explainability post hoc a complex model;
• Second, the idea of extent – global explainability methods aim to capture the logic of the entire design while local explainability only aims to capture the decision-making process of the model for a given input;
• Third, the idea of implementability – model-agnostic approaches could be applied to any kind of AI system as long as the system provides some form of input and output, while model specific approaches utilise the inner working details of a particular AI model. 

These categories help explain the different approaches used in XAI. Two of the most widely adopted post-hoc explanation techniques are SHAP and LIME. These techniques explain the prediction using distinct strategies: SHAP generates explanations by considering the importance of each feature in the final prediction, whereas LIME rationalises the non-linear model by surrogating a local model to explain its predictions (Salih et al., 2024). For example, SHAP can provide information on the purpose and weight of each feature for all instances as well as one instance (ibid.), referred to as global and local explanations respectively. An instance is an individual data point. On the other hand, LIME can only provide local explanations consequent to its architecture: that is, fitting a linear model locally to explain one single instance from the complex model. SHAP is more advantageous than LIME in terms of features such as global explanations and consistency (Roshinta, 2024) in spite of its processing load. In contrast, LIME is considerably faster and more cost-effective. These unique advantages of both post-hoc techniques demonstrate transparency and support users in interpreting complex AI models, thereby increasing trust and informed decision-making.

Real-life examples of these methods include a study conducted on diabetic patients, where a model-agnostic approach was used. The AI models were trained using explainable techniques. In this case, both LIME and SHAP provide explanations for the predictions of the patients (Ahmed et al., 2025). In another study, researchers used LIME to give reasoning to the prediction of Alzheimer’s patients of whether it is in the early stages or later (Hamza et al., 2021; Vimbi et al., 2024). Hence, this demonstrates the many ways XAI can be experimented with to improve AI systems that benefit consumers, while allowing reliability, transparency and trust.

3. Applications of XAI

3.1 Applications of XAI in Healthcare

Healthcare is one of the most important application areas for XAI because clinical decisions directly affect patient safety and wellbeing. However, in today’s healthcare environment, utilisation of non-transparent and poorly interpretable algorithms could, in the worst-case scenario, create diagnostic challenges. According to Tjoa and Guan (2021), both doctors’ willingness to adopt an AI tool regularly and their compliance are related to how the final diagnosis has been derived. Without clear explanations, clinicians may find it more difficult to recognise misleading outputs or evaluate the reliability of AI recommendations. 

One practical example can be found in medical imaging. According to Amann et al. (2020), an oncologist has significantly more confidence in a tumour grade determined by an algorithm when he or she can see which pixels in the image were used for the classification than when the decision is based solely on a black-box process. As a result, clinicians can compare these highlighted regions with their own expertise. By having visual tracing, the clinicians are prompted to systematically check and question the AI’s output: the algorithm logic can directly be compared to doctors’ experience and knowledge. This interaction supports holistic and informed decision-making rather than completely relinquishing to an automated system (Amann et al., 2020). Besides accurately diagnosing the interpretability of medical results by explaining the rationale behind each model to doctors, it facilitates better patient consulting as well as justification toward ethics boards of hospitals (Markus et al., 2021).

Beyond diagnostic imaging, XAI is also applied in early warning and risk prediction systems that support clinicians in fast-moving care settings. Lauritsen et al. (2020) developed an explainable AI early warning score (xAI-EWS) that predicts acute critical illness from electronic health records. Rather than only outputting a risk score, the system also shows which specific data points, such as recent vital signs or lab results, influenced the prediction. This allows clinicians to check the reasoning behind an alert instead of simply acting on a number they cannot question. This kind of transparency is especially important in time sensitive situations, where a wrong or unclear prediction could delay proper treatment. If a clinician can see that a warning is based on a genuine and recognisable pattern of patient deterioration, rather than a random or unclear signal, they are more likely to trust the alert and respond appropriately (Lauritsen et al., 2020).

More broadly, Vellido (2020) argues that interpretability and visualisation are not simply added extras in medical ML; they are a requirement for real-world adoption. Since healthcare decisions directly affect patient safety, models that cannot show their reasoning are harder to trust, harder to correct when they are wrong and harder to justify to patients and colleagues. Interpretability, therefore, is not only about improving accuracy or usability; it also acts as a safeguard, allowing clinicians to catch errors, question unclear predictions and stay accountable for the final decision, even as AI tools take on a bigger role in patient care.

3.2 Applications of XAI in Retail and E-Commerce

A significant application of XAI in retail and e-commerce is its contribution to the development of trustworthy AI, which promotes transparency, accountability and fairness in AI-driven decision-making. The European Commission’s High-Level Expert Group on AI (2019) identifies transparency as one of the seven fundamental requirements for trustworthy AI, emphasising that AI systems should provide explanations that enable stakeholders to understand and appropriately challenge algorithmic outcomes. In retail environments, XAI facilitates the interpretation of product recommendations, pricing strategies and automated decisions, thereby enabling consumers and businesses to assess the rationale underlying AI outputs. This demonstrates that explainability is a critical mechanism for supporting AI systems to operate in accordance with ethical principles while strengthening user confidence and encouraging responsible AI adoption.

Personalised product recommendation is the most studied application of XAI in e-commerce, since justifying the reason why a specific item was suggested to a specific user is central to building consumer trust (Zhang & Chen, 2020). To address this, Zhang et al. (2014) propose the Explicit Factor Model (EFM), generating a recommendation accompanied by a simple explanation of why, based on analysing customer reviews for positive or negative phrases. They then test this model through live experiments on the e-commerce platform JD.com. This suggests that explanations based on the language customers already use to describe products are more persuasive than justifications based on abstract internal elements of the model. However, explanations may not always be fully understood by all consumers, particularly those unfamiliar with AI systems, and overly simple explanations may omit important details about how decisions are made (Liao et al., 2021).

XAI can help address this issue by providing clearer, human-interpretable explanations alongside price recommendations and predictions (Atreya et al., 2022). This can help retailers improve transparency and accountability by allowing both businesses and customers to better understand AI-generated pricing decisions. According to Harshitha et al. (2025), consumers value perceived fairness highly. XAI can improve perceptions of fairness by supplying understandable explanations to customers, which may help build trust and loyalty between consumers and retailers. However, explainability does not guarantee that pricing decisions are fair or unbiased. However, interpretability or explainability fits the context better than comprehensibility. For instance, explanations may become overly simplified and fail to answer customers’ “why?” and “how?” questions (Liao et al., 2021). In addition, businesses may not always be able to provide complete explanations because doing so could reveal commercially sensitive pricing strategies or proprietary algorithms. Another limitation is the concept of fairwashing, which gives customers the impression that a ML model is fair while underlying biases may still exist (Aïvodji et al., 2019). These limitations suggest that explainability alone is insufficient to guarantee fairness or trustworthiness and should instead be considered as one component of responsible AI.

Overall, XAI continues to play an important role in retail and e-commerce by increasing transparency in personalised recommendation systems and dynamic pricing. By helping customers and businesses better understand AI-generated decisions, XAI can improve trust and perceptions of fairness, supporting the responsible adoption of AI. However, its effectiveness depends on the quality and clarity of the explanations provided. Explainability should be viewed as a tool that supports responsible AI rather than one that guarantees fairness or eliminates bias.

3.3 Applications of XAI in Autonomous Vehicles

AI advancements have enabled autonomous vehicles (AVs) to navigate complex environments independently. As AVs rely on complex models for real-time driving, transparent decision-making is critical, making XAI essential for trustworthy perception and control.

One application of XAI in AVs is improving the transparency of perception systems by combining sensor fusion with AI systems. AVs rely on AI-based perception systems that combine information from sensors such as LiDAR, RADAR and cameras to understand their surroundings (Atakishiyev et al., 2024). However, because these deep DL models operate as complex black boxes, XAI techniques such as SHAP and LIME are used to explain how the system interpreted sensor inputs and why particular objects or hazards were identified (Tahir et al., 2024).

Moreover,  XAI enhances the transparency of real-time decision-making by allowing AVs to analyse the surroundings and make a decision. After interpreting the input from the sensors, AI determines the best action that ensures safety for passengers. Understanding these decisions is essential for developers to improve AV systems. XAI explains why the AI selects a particular response to the stimulus such as braking or adjusting speed. Techniques such as SHAP and LIME identify which sensor inputs and model features that contributed most to a particular driving decision, improving the transparency of AI-generated actions (Kuznietsov et al., 2024).

When an AV crashes, determining the cause of a fault is difficult because AI decision logic is opaque. XAI supports this by generating readable logs of what the vehicle perceived and why it acted. These logs provide clear explanations of vehicle behaviour that can be understood by investigators, insurance adjusters and lawyers. (SmythOS, 2026). Also, transportation regulators use such records to determine whether an accident resulted from a fault in the vehicle’s autonomous system or from other road conditions (Atakishiyev et al., 2021). This positions XAI as a “black-box recorder” for AVs, making it an indispensable mechanism for assigning liability between manufacturers, software developers and human drivers.

A second barrier to AV adoption is passenger unease with unclarified machine decisions. Empirical studies have shown that providing explanations and perceptible systems significantly increases users’ trust in a system (Atakishiyev et al., 2021). Transparent decision-making allows passengers to better understand the AV’s actions, leading to increased confidence (Peintner et al., 2025). This indicates that XAI functions are not only technical add-ons, but a core safety-validation tool, since trust once damaged by unexplained failures is hard to rebuild, making transparency a design priority rather than an afterthought.

In conclusion, XAI is essential for the safe deployment of AVs. By making perception, decision-making and accident investigation transparent, it builds accountability among developers and passengers alike. As AV technology advances, embedding explainability from the start will remain key to earning public trust and ensuring safety. However, XAI may not be able to generate unique explanations in real time to stakeholders as they would need different types of explanations.

4. Ethical Considerations in XAI

While XAI improves understanding of how AI systems make decisions, it also raises ethical concerns regarding fairness, privacy, accountability and responsible AI development. One major concern is fairness and bias. AI models learn from historical data, and if that data contains existing social or demographic biases, the model can repeat or even worsen these biases in its decisions (Mehrabi et al., 2021). Explainability plays a role here because it allows developers and auditors to check whether a model’s decision is based on relevant factors or on a biased pattern of data, such as unfairly linking outcomes to race, gender or income group. Without this kind of checking, biased systems can continue to cause harm while appearing accurate at first glance.

Moreover, privacy and data protection are also central ethical issues. Many AI systems are trained on personal or sensitive data, and explanations generated by these systems can sometimes reveal more about an individual than intended. Regulatory frameworks such as the GDPR’s “right to explanation” require that individuals be given a meaningful account of automated decisions that affect them, which places a legal duty on institutions to balance transparency with data protection (Goodman & Flaxman, 2017).

Transparency and accountability are closely linked ideas. Transparency refers to how openly a system’s logic and data can be inspected, while accountability refers to who is responsible when an AI system causes harm or makes a mistake. Diakopoulos (2016) argues that clear explanations are necessary for holding both the system and the people who built or deployed it accountable, since without insight into how a decision was reached, it becomes very difficult to assign responsibility.

Human trust and human-AI collaboration are shaped directly by how well an AI system can explain itself. Shin (2021) found that explainability has a direct effect on how much users trust an AI system and how willing they are to accept its outputs. When people understand why a system produced a certain result, they are more likely to work alongside it effectively rather than blindly accepting or completely rejecting its suggestions.

Finally, AI regulations and responsible AI development tie all these considerations together. Floridi et al. (2018) present an ethical framework built around principles such as fairness, transparency and accountability, arguing that these should guide AI development from the start rather than being added afterward. As regulations continue to develop worldwide, explainability is increasingly treated not as an optional feature, but as a basic requirement for AI systems to be considered trustworthy, fair and safe for real-world use.

5. The Future of XAI

As AI moves toward higher levels of integration across critical domains like healthcare, finance and transportation, the future of XAI is likely to be defined by greater standardisation and resolving the trade-off between model performance and interpretability. Currently, safety and regulatory standards remain underdeveloped. While regulators are working on frameworks like ISO/AWI PAS 8800 for vehicle safety and AI, its scope remains in a developmental phase (Kuznietsov et al., 2024). Standardised guidance is critical because clear benchmarks ensure consistency across developers, uphold physical safety and foster public confidence in high-stakes AI. At the same time, research faces a persistent tension between interpretability and performance. This dilemma is pushing developers toward hybrid architectures combining transparent, rule-based reasoning with deep neural networks or inherently transparent AI designs (IRE Journals, 2025). This trajectory suggests that the next phase of XAI may not simply add explanation layers onto black-box models, but could restructure how systems are built, embedding interpretability directly into decision-making pipelines.

Beyond regulation and architecture, XAI is also expected to be shaped by advances in user-centred explanation delivery. For instance, researchers in autonomous driving have suggested the SafeX framework, a system designed to format and present complex driving decisions to human users in real time, enabling explanation delivery while ensuring model safety (Kuznietsov et al., 2024). Likewise, work on capsule network architectures that group neurons to better model spatial relationships suggests perception systems could become naturally understandable, recognising objects in varied poses with built-in explanatory value rather than relying on post-hoc techniques (Atakishiyev et al., 2021).

However, these prospective developments face open research challenges. It remains uncertain whether inherently interpretable models can achieve performance parity with black-box systems, how explanation quality can be measured objectively or how transparency might expose security vulnerabilities. Furthermore, adapting explanations to diverse stakeholders from domain experts to end users presents an ongoing design hurdle.

Ultimately, while AVs illustrate these core trends, demands for user-centred explanations, benchmarking and integration into responsible governance frameworks are similarly transforming healthcare, finance and generative AI. As AI systems become increasingly integrated into society, future research must ensure that explainability evolves alongside advances in performance, safety and ethical governance.

6. Conclusion

AI has transformed the way decisions are made across many sectors, but as these systems become increasingly complex, understanding how these decisions are made is crucial. The advancement of XAI can help solve the issue of explainability in complex models. XAI assists complex models, rather than replacing them, by making them more transparent and accountable, ensuring that their decisions can be interpreted and understood by users ranging from developers to everyday users.

The importance of XAI extends to high-demand industries such as healthcare, retail, commerce and AV production. Without transparent reasoning, AI systems may struggle to gain acceptance and reliance across varying users. As a result, XAI proves to be an integral component in today’s systems and has become the definitive framework for bridging the gap between users and AI systems.

However, the explainability itself is not perfect and has some drawbacks. Achieving a balance between explainability, accuracy and efficiency, along with cost, remains a major area of ongoing research, particularly when AI models become more complex. Along with its challenges, XAI suffers from ethical concerns regarding bias and fairness in its explainability, accountability and data privacy. Addressing these challenges will be essential in achieving the true potential of XAI.

Looking ahead, the future of XAI is still growing and will continue to evolve and expand in many complex systems. Continued research can help improve scalability, transparency and address the ethical considerations that can help make XAI more versatile and more applicable in the real world. Most importantly, XAI’s explainability can help develop AI systems whose decisions can be understood, justified and trusted, enabling developers and common users to embrace the benefits of AI.

Ahmed, S., Kaiser, M.S., Hossain, M.S. and Andersson, K. (2024) A Comparative Analysis of LIME and SHAP Interpreters with Explainable ML-Based Diabetes Predictions, IEEE Access, 13, pp. 37370-37388.

Aïvodji, U., Arai, H., Fortineau, O., Gambs, S., Hara, S. and Tapp, A. (2019) Fairwashing: The risk of rationalization, arXiv: 1901.09749. <https://arxiv.org/abs/1901.09749>

Arrieta, A.B., Díaz-Rodríguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado, A., Garcia, S., Gil-Lopez, S., Molina, D., Benjamins, R., Chatila, R. and Herrera, F. (2020) Explainable Artificial Intelligence (XAI): Concepts, taxonomies, Opportunities and Challenges toward Responsible AI, Information Fusion, 58(1), pp. 82-115.

Atakishiyev, S., Salameh, M., Yao, H. and Goebel, R. (2021) Explainable artificial intelligence for autonomous driving: A comprehensive overview and field guide for future research directions, arXiv: 2112.11561. <https://arxiv.org/abs/2112.11561>

Atakishiyev, S., Salameh, M., Yao, H. and Goebel, R. (2024) Explainable artificial intelligence for autonomous driving: A comprehensive overview and field guide for future research directions, IEEE Access, 12, pp. 101603-101625.

Atreya, M., Chhibber, N. and Singh, H. (2022) Explainable Machine Learning for Dynamic Pricing in Fast-Changing Retail Environments, SSRN Electronic Journal.

Barredo Arrieta, A., Díaz-Rodríguez, N., Del Ser, J., Bennetot, A., Tabik, S., Barbado, A., García, S., Gil-López, S., Molina, D., Benjamins, R., Chatila, R. and Herrera, F. (2020) Explainable Artificial Intelligence (XAI): Concepts, taxonomies, opportunities and challenges toward responsible AI, Information Fusion, 58, pp. 82-115.

Diakopoulos, N. (2016) Accountability in algorithmic decision making, Communications of the ACM, 59(2), pp. 56-62.

Dwivedi, R., Dave, D., Naik, H., et al. (2023) Explainable AI (XAI): Core Ideas, Techniques, and Solutions, ACM Computing Surveys, 55(9), Article 194, pp. 1-33.

European Commission High-Level Expert Group on Artificial Intelligence (2019) Ethics Guidelines for Trustworthy AI (Luxembourg: Publications Office of the European Union). <https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-ai>

Floridi, L., Cowls, J., Beltrametti, M., Chatila, R., Chazerand, P., Dignum, V., Luetge, C., Madelin, R., Pagallo, U., Rossi, F., Schafer, B., Valcke, P. and Vayena, E. (2018) AI4People—An ethical framework for a good AI society: Opportunities, risks, principles, and recommendations, Minds and Machines, 28(4), pp. 689-707.

Goodman, B. and Flaxman, S. (2017) European Union regulations on algorithmic decision-making and a “right to explanation”, AI Magazine, 38(3), pp. 50-57.

Guidotti, R., Monreale, A., Ruggieri, S., Turini, F., Giannotti, F. and Pedreschi, D. (2018) A survey of methods for explaining black box models, ACM Computing Surveys, 51(5), Article 93, pp. 1-42.

Gunning, D. and Aha, D. (2019) DARPA’s Explainable Artificial Intelligence (XAI) Program, AI Magazine, 40(2), pp. 44-58.

Harshitha, N.S., Sampathkumari, V., G, R. and C, P. (2025) AI in Dynamic Pricing: How Fairness and Trust Influence Customer Loyalty in E-commerce, International Journal on Economic and Business Management, Volume XIV, 48(1), pp. 326-331.

Huang, J., Wang, Z., Li, D. and Liu, Y. (2022) The Analysis and Development of an XAI Process on Feature Contribution Explanation, 2022 IEEE International Conference on Big Data (Big Data), pp. 5039-5048.

Jordan, M. and Mitchell, T. (2015) Machine learning: Trends, perspectives, and prospects, Science, 349(6245), pp. 255-260.

Kalasampath, K., Spoorthi, K.N., Sajeev, S., Kuppa, S.S., Ajay, K. and Maruthamuthu, A. (2025) A Literature Review on Applications of Explainable Artificial Intelligence (XAI), IEEE Access, 13, pp. 41111-41140.

Kaya, O., Shah, A.F.M.S., Karabulut, M.A., Karahan, S.N., Osmanca, M.S. and Acir, N. (2026) Explainable Artificial Intelligence (XAI): Concepts, Applications, Challenges, and Future Perspectives, IEEE Access, 14, pp. 27394–27417.

Knap, M. (2022) Model-Agnostic XAI Models: Benefits, Limitations and Research Directions, Bachelor’s thesis. Delft University of Technology. <https://repository.tudelft.nl/file/File_665977bd-f748-492a-b820-fffdc52ada3f?preview=1>

Kuznietsov, A., Gyevnar, B., Wang, C., Peters, S. and Albrecht, S.V. (2024) Explainable AI for safe and trustworthy autonomous driving: A systematic review, IEEE Transactions on Intelligent Transportation Systems, 25(12), pp. 19342-19364.

LeCun, Y., Bengio, Y. and Hinton, G. (2015) Deep learning, Nature, 521(7553), pp. 436-444.

Liao, Q., Gruen, D., Miller, S., Smith, D., et al. (2021) Question-Driven Design Process for Explainable AI User Experiences, Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, 275, pp. 1-15.

McCarthy, J. (2007) From here to human-level AI, Artificial Intelligence, 171(18), pp. 1174-1182.

McCarthy, J., Minsky, M., Rochester, N. and Shannon, C. (1955) A proposal for the Dartmouth summer research project on artificial intelligence (Dartmouth College).

Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K. and Galstyan, A. (2021) A survey on bias and fairness in machine learning, ACM Computing Surveys, 54(6), pp. 1-35.

Mersha, M., Lam, K., Wood, J., AlShami, A.K. and Kalita, J. (2024) Explainable artificial intelligence: A survey of needs, techniques, applications, and future direction, Neurocomputing, 599, Article 128111.

Novelli, C., Taddeo, M. and Floridi, L. (2023) Accountability in artificial intelligence: what it is and how it works, AI & Society, 39(1).

Omeiza, D., Webb, H., Jirotka, M. and Kunze, L. (2021) Towards accountability: Providing intelligible explanations in autonomous driving, Proceedings of the 2021 IEEE Intelligent Vehicles Symposium (IV), pp. 685-692.

Peintner, J., Manger, C., Berger, M. and Riener, A. (2025) Balancing transparency and control: The impact of AI explanation detail on user perception in automated vehicles, Transportation Research Interdisciplinary Perspectives, 34, Article 101742.

Phillips, P.J., Hahn, C.A., Fontana, P.C., Yates, A.N., Greene, K.K., Broniatowski, D.A. and Przybocki, M.A. (2021) Four Principles of Explainable Artificial Intelligence (Gaithersburg, MD: National Institute of Standards and Technology).

Roshinta, T.A. and Gábor, S. (2024) A Comparative Study of LIME and SHAP for Enhancing Trustworthiness and Efficiency in Explainable AI Systems, 2024 IEEE International Conference on Computing (ICOCO), pp. 134-139.

Rudin, C. (2019) Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead, Nature Machine Intelligence, 1(5), pp. 206-215.

Salih, A.M., Raisi-Estabragh, Z., Boscolo Galazzo, I., Radeva, P., Petersen, S.E., Menegaz, G. and Lekadir, K. (2023) A Perspective on Explainable Artificial Intelligence Methods: SHAP and LIME, Advanced Intelligent Systems, 7(1).

Shad, H.A., Rahman, Q.A., Asad, N.B., Bakshi, A.Z., Mursalin, S.F., Reza, M.T. and Parvez, M.Z. (2021) Exploring Alzheimer’s disease prediction with XAI in various neural network models, TENCON 2021–2021 IEEE region 10 conference (TENCON), pp. 720-725.

Shin, D. (2021) The effects of explainability and causability on perception, trust, and acceptance: Implications for explainable AI, International Journal of Human-Computer Studies, 146, Article 102551.

Tahir, H.A., Alayed, W., Hassan, W.U. and Haider, A. (2024) A novel hybrid XAI solution for autonomous vehicles: Real-time interpretability through LIME–SHAP integration, Sensors, 24(21), Article 6776.

Turing, A. (1950) Computing machinery and intelligence, Mind, 59(236), pp. 433-460.

Ugboko, R. and Oloruntoba, O. (2025) Explainable Artificial Intelligence in Autonomous Vehicles: Methodologies, Challenges, and Prospective Directions, Iconic Research and Engineering Journals, 8(10), pp. 1578-1593.

Vimbi, V., Shaffi, N. and Mahmud, M. (2024) Interpreting artificial intelligence models: a systematic review on the application of LIME and SHAP in Alzheimer’s disease detection, Brain Informatics, 11(1), Article 10.

Zhang, Y. and Chen, X. (2020) Explainable Recommendation: A Survey and New Perspectives, Foundations and Trends in Information Retrieval, 14(1), pp. 1-101.

Zhang, Y., Lai, G., Zhang, M., Zhang, Y., Liu, Y. and Ma, S. (2014) Explicit Factor Models for Explainable Recommendation Based on Phrase-Level Sentiment Analysis, Proceedings of the 37th International ACM SIGIR Conference on Research and Development in Information Retrieval, pp. 83-92.