Toward Socially Accountable Data Science Education: Proposing a Conceptual Framework for Integrating Explainable AI (XAI) and Accountability Principles

1. Introduction

When an algorithm denies someone a loan, flags a job applicant for rejection, or assigns a risk score that impacts a criminal’s sentence, the consequences are immediate and have real human consequences. Yet, the person affected may never know why the decision was made, who bears responsibility for it, or how it might be challenged. These are not hypothetical scenarios. They are documented patterns across high-stakes domains including healthcare, employment, credit, education, and criminal justice – domains in which AI systems now routinely mediate access to opportunity and resources (O’Neil, 2016; Eubanks, 2018). These patterns are corroborated by a growing body of peer-reviewed empirical scholarship documenting algorithmic bias and discrimination across healthcare, employment, and legal contexts (Chen et al., 2023; Wang et al., 2024; Chen, 2023; Hughes et al., 2025). The persistence of these harms should not be misperceived as solely a technical problem. It is, in significant part, an educational one. The practitioners who design and deploy these systems are being trained in programs that have historically prioritized predictive performance, computational efficiency, and algorithmic optimization, while treating ethical considerations as supplementary content rather than foundational competencies (Saltz et al., 2018). A recent review of undergraduate data science education confirms that ethics training remains underdeveloped across programs (Dogucu et al., 2025). The result is a growing workforce of technically sophisticated practitioners who are not adequately prepared to reckon with the social consequences of the systems they build. This paper proposes a framework for closing that gap.

This paper proposes a solution to this gap in data science education by offering a pedagogical framework that integrates explainable AI (XAI) and social accountability as core competencies within higher education programs that train AI developers. Throughout this paper, the term “AI developer” refers to students in data science, computer science, and information science programs who are learning to design, build, and deploy AI systems — individuals whose technical decisions will carry direct social consequences upon entering professional practice. This proposed framework is grounded in the belief that technical decisions also carry critical social implications. Choices made about data, model architecture, evaluation metrics, and deployment contexts reflect and reproduce values, priorities, and power relations (Maas, 2023). In order to prepare future practitioners to navigate these dimensions, it is necessary to provide more than mere exposure to ethical principles. It is necessary to encourage the growth and support of concrete habits of transparency, answerability, and reflexivity that can be enacted throughout the AI development lifecycle.

This framework for data science education is organized around four pillars of social accountability:

• Answerability, the obligation to explain and justify decisions to those affected by them;
• Responsibility, the ongoing commitment to anticipate and address harm across the system lifecycle;
• Enforcement, the institutional and regulatory mechanisms that give accountability operational force;
• Reflexivity, the critical examination of the assumptions, biases, and values that developers bring to their work.

This framework is developed from the perspective of higher education programs in data science, computer science, and information science at the undergraduate and graduate levels, where students are being trained to design, build, and deploy machine learning and AI systems in professional contexts. The principles and practices described here are most immediately applicable to courses in machine learning, data science ethics, applied AI, and capstone or project-based courses where students work on systems intended for real-world deployment. However, the framework is designed to be adaptable across program types and institutional contexts, and educators in adjacent fields, including statistics, information science, and software engineering, will find relevant applications throughout.

These pillars are further situated within four analytical dimensions – technical, social, organizational, and political – that together highlight the multidimensional nature of accountable AI development. By embedding explainability and accountability at multiple levels of data science education, this framework aims to contribute to the formation of practitioners who understand that the value of an AI system cannot be measured by accuracy alone. Systems that are powerful but opaque, or efficient but unjust, fail in their fundamental purpose. The goal of this work is to help ensure that the next generation of data scientists is equipped with the knowledge and skills to build AI systems that are worthy of public trust.

2. Materials and Methods

As a conceptual paper rather than an empirical study, this section describes the approach for synthesizing the literature used to produce the proposed framework. This framework was developed through a narrative synthesis rather than a systematic review. The synthesis followed the interpretive, argument-building logic characteristic of conceptual and theoretical papers in this field, drawing on three literatures: (1) accountability theory as articulated in public administration and governance scholarship, particularly Bovens’ (2007) conceptualization of answerability and the distinction between processes of explanation and processes of consequence; (2) the explainable AI (XAI) literature, including technique-level scholarship on interpretability methods (e.g., LIME, SHAP, counterfactual explanations, model cards) and critical scholarship questioning the sufficiency of explanation alone as a guarantor of accountability (Mittelstadt et al., 2019; Rudin, 2019); and (3) data science and information science education scholarship addressing the integration of ethics, reflexivity, and governance into curricula (Davis, 2020; Saltz et al., 2018; Dogucu et al., 2025).

Sources were identified through iterative search of major databases (Web of Science, Scopus, Google Scholar) using combinations of terms including “explainable AI,” “algorithmic accountability,” “AI ethics education,” “data science curriculum,” and “responsible AI governance,” along with citation chaining from key theoretical works (e.g., Bovens, 2007; Diakopoulos, 2016; Doshi-Velez and Kim, 2017). Sources were prioritized for inclusion based on their direct relevance to accountability theory, XAI methods, or data science pedagogy. This search and prioritization process was conducted iteratively by the authors rather than through a fixed date range, formal inclusion/exclusion criteria, or independent dual-reviewer screening, with the goal being conceptual coverage and argumentative coherence rather than exhaustive or reproducible retrieval.

The four-pillar structure (answerability, responsibility, enforcement, reflexivity) was derived analytically from accountability theory (Bovens, 2007) and adapted to the context of AI development education through iterative synthesis with the XAI and pedagogy literatures described above. The four analytical dimensions (technical, social, organizational, political) are an original typology developed for this paper rather than one adopted wholesale from an existing framework, though it draws on the levels-of-analysis logic common in sociotechnical systems scholarship, which routinely distinguishes artifact-level, interactional, institutional, and macro-structural levels of explanation (e.g., Diakopoulos, 2016; Ananny and Crawford, 2018). These four dimensions were selected, because each corresponds to a distinct locus of accountability failure identified in the XAI literature reviewed for this paper: technical failures in interpretability and fidelity, social failures in trust and comprehension, organizational failures in oversight and incentive structures, and political failures in enforcement and power asymmetry. The illustrative scenario presented in Section 4.1 is a hypothetical, composite case constructed by the authors to demonstrate how the four pillars operate jointly within a single educational project.

3. Results

3.1. Answerability

Answerability is the responsibility of an individual or model to provide explanations and justifications for their decisions to those affected by these decisions (Bovens, 2007). It is a significant component of accountability, or the duty to give reasons before any judgment or sanction is imposed. Within the context of artificial intelligence development, answerability requires that developers and institutions be able to explain how algorithmic systems reach their decisions, what data they rely on to inform these decisions, and on what grounds those decisions can be evaluated. As AI systems increasingly control access to information, economic opportunity, and civic participation, teaching aspects of answerability is critical for fostering social accountability in AI development.

3.1.1. Algorithmic Systems Complicate Traditional Answerability

Machine learning models typically utilize complex, and in many cases unknowable, statistical processes that make a straightforward interpretation challenging. Burrell (2016) identifies multiple forms of opacity in machine learning systems, including technical complexity and proprietary secrecy, both of which limit meaningful explanation. Similarly, Diakopoulos (2016) argues that algorithmic accountability depends not merely on transparency but on the ability to articulate how decisions are made and how they may be questioned. Without answerability, affected individuals may experience consequences without access to reasons for the decisions made.

These challenges are fundamental issues in both data and information science. Historically, these fields have concerned themselves with the organization, retrieval, and mediation of knowledge (Saracevic, 1999). AI systems now perform these mediating functions at a previously unimaginable scale, determining which information is surfaced, prioritized, or acted upon. Floridi (2011) conceptualizes this environment as the “infosphere,” whereby informational agents shape human agency and social reality. AI developers, therefore, must be aware of their role not only as engineers but also as stewards of information infrastructure. Their systems structure epistemic access and influence social outcomes. Teaching answerability is therefore intrinsic to the ethical and professional formation of AI practitioners.

To embed answerability into AI education, it is necessary to move beyond abstract ethical principles toward concrete ethical AI development practices. First, curricula should incorporate interpretability techniques that enable students to operationalize explanation in technical systems (Doshi-Velez and Kim, 2017). Understanding feature importance, model behavior, and uncertainty estimation prepares future developers with the skills to design systems that can justify their outputs. Second, structured documentation practices, such as model cards and data sheets, should be integrated as core components of coursework. These artifacts cultivate habits of transparency, traceability, and reflection. Third, case-based examination of documented algorithmic harms, such as those described by O’Neil (2016), situates technical design within real-world social consequences, reinforcing the relational nature of accountability.

Answerability must be framed as a professional necessity rather than as a regulatory burden in order to successfully be integrated into the learning of students. When students internalize the expectation that AI systems must be explainable to affected communities, policymakers, and interdisciplinary stakeholders, they develop reflexive awareness of the broader implications of their work (Abgrall et al., 2024). Social accountability in AI development does not begin with legislation alone; it begins in the classroom. By integrating conceptual clarity, technical competence, and normative reflection, information science education can help ensure that AI systems are not only powerful and efficient but also justifiable and responsive to the societies they shape.

3.2. Responsibility

In the modern day, systems play a major role in influencing decisions about loans, medical diagnoses, hiring, and education. Because these systems impact real-world opportunities, social accountability must be a key aspect of developers’ training (Rashid and Kausik, 2024). Responsibility is not a one-time step in development but rather an ongoing commitment across a system’s lifecycle, from its design and data selection to deployment and review. This means thinking ahead about possible harm rather than defending poor decisions later. AI systems are not neutral (González Sendino et al., 2024). The impact of these systems reflects the choices people make. Responsibility rests with those who build and manage them.

3.2.1. Responsibility as a Shared but Unequal Obligation

Responsibility is shared among individuals, organizations, and society, but it is not equally distributed among these different groups. Greater authority brings a greater amount of accountability. Developers and researchers must question assumptions, examine data carefully, and weigh trade-offs between accuracy and fairness. If a medical risk model relies on historical healthcare spending to estimate need, a responsible developer should ask whether that spending reflects illness or unequal access to care (Obermeyer et al., 2019). Raising such concerns before deployment is part of professional duty.

Institutions must ensure that systems operate responsibly on a scale. This includes governance, audits, transparency, and regular review. If a hiring system favors candidates from narrow backgrounds, a responsible organization reassesses it and adds safeguards rather than focusing solely on efficiency (Raghavan et al., 2020). Governments and communities must decide where these systems should and should not be used. When cities restrict the use of predictive policing tools because certain neighborhoods are disproportionately targeted, they make it clear that efficiency cannot outweigh fairness (IEEE, 2019). Public oversight, clear rules, and independent review help ensure that powerful systems reflect shared values rather than operating without limits.

3.2.2. Teaching Responsibility as Practice

Responsibility should be taught as a professional habit, not a brief ethics lesson tacked onto a longer lecture. Students need to learn to ask who benefits from a system, who might be harmed, and what protections are necessary. They should also understand that responsibility continues after deployment. Monitoring results, responding to unintended effects, and revising systems as needed are part of responsible work (Lungu and Tabur, 2025). When this mindset is reinforced throughout training, it becomes part of professional identity rather than an afterthought.

Responsibility in socially accountable development requires deliberate action. The way forward requires that individuals must think critically, organizations must embed oversight in their processes, and societies must set clear limits. When responsibility is treated as central, systems are more likely to support fairness, safety, and public trust. Preparing future professionals with this perspective helps ensure that innovation strengthens the common good rather than deepening inequality.

3.3. Enforcement

3.3.1. Rationale for Enforcement

While significant progress has been made in describing the principles of explainability and responsible AI, a remaining challenge in the literature is the lack of effective enforcement approaches to ensure that developers implement these principles in their design. Studies of global AI ethics frameworks show that most of these initiatives emphasize normative values such as transparency, fairness, and accountability, but often offer limited mechanisms for operational oversight or compliance (Jobin et al., 2019; Camilleri, 2024). The gap between principle and practice is not incidental. Instead, this reflects a structural failure to build consequences into accountability frameworks. Diakopoulos (2016) argues that algorithmic systems should be subject to public accountability, particularly when their decisions shape access to information, services, or opportunities with social consequences.

3.3.2. Enforcement Deficits in Data Science Education

Within data science education, this gap in enforcement of ethical principles is particularly significant. Educational programs play a critical role in developing the knowledge and practices of future AI developers. However, enforcement-oriented governance mechanisms are rarely included in curricula of these programs. Although many institutions now incorporate ethics modules into AI and data science programs, these often focus on conceptual awareness of these principles rather than on compliance structures or accountability processes. Research on AI governance highlights the critical need to operationalize accountability through institutional practices, oversight mechanisms, and compliance processes that translate ethical principles into practice (Novelli et al., 2023). Without these mechanisms, students may develop advanced technical skills without really understanding how explainability and social responsibility are evaluated, audited, or enforced in real-world environments. Therefore, enforcement in education requires tying explainability to evaluation standards and academic consequences, including structured review, documentation requirements, and assessment criteria that reward responsible design. Such mechanisms align with the judgment and consequence components of accountability frameworks, which distinguish meaningful accountability from symbolic transparency.

3.3.3. Structural and Sociotechnical Challenges

A further complication arises from the distributed nature of responsibility in sociotechnical AI systems. Development and deployment typically involve multiple parties, including developers, institutions, regulators, and end users. The diffusion of responsibility among these parties complicates enforcement because accountability may become fragmented or unclear. Research on algorithmic accountability consistently stresses the need for clearly defined oversight structures, reporting and documentation practices, and institutional accountability arrangements that extend beyond individual developers (Diakopoulos, 2016; Novelli et al., 2023). In educational environments, this suggests that accountability should not be treated solely as an individual ethical disposition but as a structured process supported by governance mechanisms. Without clearly defined forums for review and evaluation, even well-articulated explainability frameworks may fail to produce consistent compliance.

Technical limitations also make enforcement challenging. Many advanced machine learning systems still pose interpretability challenges that make auditing and verification difficult. Despite significant advances in explainable AI research, important challenges remain regarding the fidelity, stability, and human interpretability of explanations (Adadi and Berrada, 2018). These constraints limit institutions’ and regulators’ ability to define uniform, enforceable transparency standards. Enforcement frameworks must therefore recognize the evolving state of interpretability research and incorporate mechanisms for iterative evaluation rather than static compliance checklists.

3.3.4. A Multi-Level Enforcement Framework

To address these challenges, this paper offers a multi-level enforcement framework for data science education. The first layer of this framework comprises legal and policy enforcement, in which external regulatory frameworks, professional standards, and public governance mechanisms set baseline expectations for transparency and accountability. These frameworks create formal consequences, such as regulatory investigations or funding restrictions, that incentivize institutional compliance. The second layer focuses on institutional and administrative enforcement within universities. This section covers curriculum requirements, ethical review procedures, documentation standards, and grading systems that implement responsible AI principles and enforce academic consequences like mandatory revisions, project rejection, or denial of certification. The third layer emphasizes technical enforcement by integrating explainability practices into development processes through model documentation, auditing, and regular assessment of explanation quality.

3.3.5. Implementation of Constraints and Future Directions

Enforcement mechanisms for explainability and social responsibility are still evolving. Comparative analyses of AI governance highlight significant variation across jurisdictions in ethical priorities, policy approaches, and implementation strategies (Jobin et al., 2019). Technical opacity, uneven institutional capacity, and divergent policy approaches limit immediate harmonization. Consequently, effective enforcement is likely to emerge gradually through iterative policy refinement, institutional learning, and advances in interpretability research. By embedding layered mechanisms of oversight and consequence within data science education, explainability can move beyond an aspirational principle toward meaningful social accountability.

3.4. Reflexivity

As artificial intelligence systems increasingly influence decisions in domains such as healthcare, finance, and education, ensuring social accountability in AI development has become a critical concern. A key teaching focus in data science education is reflexivity. Reflexivity is the practice of critically examining one’s assumptions, biases, values, and positionality when designing, implementing, and evaluating technological systems. In AI development, reflexivity encourages future data scientists to recognize how their perspectives and decisions shape the behavior and societal impact of AI models.

Traditional data science education often focuses heavily on technical competencies such as model accuracy, optimization, and computational efficiency. While these skills are essential, they do not fully address the broader societal implications of AI systems. Reflexivity helps bridge this gap by encouraging students to reflect on the social context in which AI systems operate and how human choices shape algorithmic outcomes. Developers make numerous subjective decisions during the AI development process, including data selection, feature engineering, labeling strategies, and evaluation metrics. Each of these choices can introduce unintended biases or reinforce existing inequalities. Teaching reflexivity helps students recognize these influences and consider their ethical implications.

Recent research has emphasized the importance of reflexive practices in data science and AI development. Cambo and Gergle (2022) introduce the concept of computational reflexivity, which encourages developers to explicitly acknowledge how modeling choices and personal perspectives shape the resulting system. Similarly, Boyd (2025) highlights reflexivity as an essential component of responsible AI research, arguing that reflective practices help identify hidden assumptions embedded in data-driven systems. These perspectives suggest that reflexivity should be integrated into AI education not as an optional ethical discussion but as a core component of the development process.

In practical terms, reflexivity can be integrated into data science education through several pedagogical strategies. For example, students can be encouraged to write positionality statements that articulate their perspectives, values, and assumptions when working with datasets. Reflective assignments can also prompt students to examine how dataset selection, labeling practices, and feature choices might influence model outcomes. Additionally, tools such as model documentation frameworks, including model cards and datasheets for datasets, provide structured ways for students to reflect on and communicate the limitations and societal implications of their models.

Teaching reflexivity also complements the goals of explainable AI (XAI). While XAI techniques aim to make model decisions interpretable to users, reflexivity focuses on making the development process itself more transparent and accountable. When developers actively reflect on their assumptions and design decisions, they are better positioned to produce explanations that are meaningful, responsible, and socially aware. In this sense, reflexivity strengthens the broader framework of social accountability by ensuring that explainability extends beyond technical explanations to encompass critical reflection on the human factors that shape AI systems. In conclusion, incorporating reflexivity into data science education not only enhances technical proficiency but also fosters a socially responsible mindset among future practitioners. This critical self-examination of assumptions and societal impacts ensures the development of AI systems that are both effective and ethically grounded, promoting greater accountability in AI innovation.

3.5. Dimensions for Applying XAI for Social Accountability

3.5.1. Technical Dimensions

Explainable Artificial Intelligence (XAI) comprises methods that make AI systems more transparent, understandable, and accessible to human users (McDermid et al., 2021). Recent literature reviews confirm that the range of XAI techniques and their application domains has expanded rapidly, from classical feature-attribution methods to large language model (LLM)-based explanation generation (Kalasampath et al., 2025; Bilal et al., 2025). In the context of social accountability, these methods matter not primarily because of their technical sophistication but because of what they make possible socially. Affected individuals gain the ability to understand why a decision was made. Institutions gain the ability to audit and validate system behavior. Developers gain the ability to identify and share the limitations of their own systems before harm occurs. Techniques such as LIME and SHAP generate explanations of how input features influence individual predictions, making visible the factors driving a decision in terms that can be communicated to non-technical audiences (Ahmed et al., 2024). Structured documentation tools such as model cards record intended use cases, known limitations, and evaluation results in formats designed for policymakers, domain experts, and affected communities, as well as technical reviewers (Mitchell et al., 2019). Counterfactual explanations show users what would need to change for a different outcome to result, a particularly powerful tool for accountability because it gives affected individuals a concrete basis for understanding and potentially challenging decisions made about them. Camilleri (2026) similarly argues that operationalizing XAI requires moving from abstract interpretability principles toward concrete tools, frameworks, and organizational protocols that translate explanation into practice. The technical aspects of XAI discussed here should be understood not as ends in themselves but as instruments through which the social, organizational, and political aspects of accountability can be put into practice.

Model transparency and the degree to which a model is understandable by design indicate whether its structure, parameters, and reasoning process can be directly examined by humans (Chamola et al., 2023). Models that are understandable by design make it possible to see how predictions are formed without relying entirely on external explanation techniques to interpret them. This is a fundamental aspect of XAI because it improves trust, supports error analysis, and enables more accountable decision-making (Hassija et al., 2024). It may also enhance general AI literacy skills by increasing awareness of issues related to how AI works (Lund et al., 2026). Understanding the behavior of complex machine learning models often requires explanation techniques applied after training them. These approaches generate explanations that help users understand how predictions are produced, particularly when the model’s internal logic is difficult to interpret directly (Dwivedi et al., 2023). Methods such as LIME and SHAP reveal how input features influence individual predictions and identify the factors driving them (Ahmed et al., 2024). Such techniques improve transparency, support model validation, and help build trust in black-box systems (Chamola et al., 2023).

How a model’s reasoning is presented to the user matters. This is especially true in systems where decisions depend on multiple input sources with different structures and purposes (Rodis et al., 2024). A useful explanation should capture each source’s relative contribution as well as the relationships among them. How explanatory information is presented shapes how well users can act on it. For example, natural language, visual highlights, feature contributions, decision rules, and high-level concepts each carry different strengths depending on who is reading them and why. These approaches are not equally effective for all users or applications, as the ability to understand an explanation is shaped by the user’s background knowledge, task requirements, and cognitive constraints (Bertrand et al., 2022; Malandri et al., 2023). An explanation that is technically accurate may still fail if it is overly complex, poorly structured, or difficult to relate to the decision at hand. Accordingly, whether a user can actually understand an explanation is a central concern for XAI because it links explanation quality to real human understanding, usability, and oversight.

How accurately an explanation reflects the way a model arrived at its prediction is a critical measure of that explanation’s quality. An explanation should identify the features and relationships the model actually uses rather than merely provide a clear or appealing summary of them. This is significant because an explanation can be easy to understand while still misrepresenting the model’s actual decision-making process (Miró-Nicolau et al., 2024; Miró-Nicolau et al., 2025). Accuracy of this kind is therefore a fundamental standard in XAI, particularly when explanations are used for validation, auditing, and decision-making in high-stakes contexts. How XAI communicates the reliability of model predictions alongside their explanations is equally important, as users need to understand not just what the model decided but how confident it was in doing so (Cheng et al., 2026). Sources of uncertainty may include noisy inputs, incomplete training data, or cases that differ from those encountered during training. Sharing this information helps prevent over-reliance on AI outputs and supports more careful and well-informed human oversight of them (Chiaburu et al., 2025).

Explainability is a fundamental component of trustworthy artificial intelligence, as it enables examination of how models operate and how their decisions can be evaluated. XAI plays a critical role in assessing key attributes of trustworthy AI, including fairness, robustness, accountability, safety, and transparency (Rawal et al., 2021). Explanations serve to reveal hidden biases, identify unreliable reasoning patterns, and support auditing and human oversight of them, particularly in high-stakes areas (Marques-Silva and Ignatiev, 2022). XAI should therefore be understood not solely as a means of interpreting model output but as a fundamental mechanism for the development and assessment of trustworthy AI systems.

3.5.2. Social Dimensions

Explainability is a key ingredient in developing and maintaining trust and user acceptance among those interacting with an AI system (Hur et al., 2025). Individuals are more open to using and trusting a system that demonstrates how it arrived at a specific decision (Ali et al., 2023; Arrieta et al., 2020). In high-stakes areas such as healthcare and finance, explainability becomes even more important because outcomes are closely tied to human well-being, and a single wrong decision could have serious consequences for those affected by it (Adadi and Berrada, 2018; Hashemi et al., 2025; Longo et al., 2024). From an ethical standpoint, explainability may also help identify and reduce bias by revealing the evidence algorithms use to make decisions and the populations most affected by them (Adadi and Berrada, 2018).

There is a growing need to develop ways of thinking that help users interpret and evaluate explanations that approximate the behavior of black-box models (Hashemi et al., 2025). The General Data Protection Regulation (GDPR) has introduced a legal framework for automated decision-making that includes the “right to an explanation,” requiring organizations to provide meaningful explanations of the logic behind decisions made by algorithmic systems and the individuals subject to them (Adadi and Berrada, 2018; Amparore et al., 2021). Explainability remains a major barrier to the full-scale adoption of AI-based tools in high-stakes areas such as healthcare, where users play various roles and may not be experts in machine learning (Arrieta et al., 2020). Recent advances in the healthcare sector have achieved high accuracy in early disease detection, patient-specific dosing, and mortality risk prediction (Alowais et al., 2023). However, high predictive performance does not always translate into user acceptance of it. Studies have shown that explanations aligning with user expectations are more likely to be accepted and used than those that simply present accuracy metrics (Hur et al., 2025; Tonekaboni et al., 2019).

Workload and task context also shape how much users rely on XAI outputs, with higher cognitive load associated with greater but less discriminating reliance on system-provided explanations (Alami et al., 2025). Recent studies also show that explanation formats play a key role in driving user acceptance, trust, and decision-making behavior. Hur et al. (2025) found that verbal explanations foster greater trust and higher user acceptance, while Papenkordt et al. (2025) demonstrated greater reliance on AI-based recommendations with verbal than with numerical explanations. Further studies show that users often prefer narrative explanations over complex technical visualizations (Zytek et al., 2024). Taken together, these findings establish that explainability is no longer simply an afterthought. It actively shapes users’ patterns of trust, reliance, and decision-making.

Because explainability is so closely tied to trust, accountability, user acceptance, and reliance on an AI system, it must be integrated into model development, validation, and use rather than treated as an afterthought. This makes explainability a core professional competency for data scientists, one that should be studied alongside the technical work of developing and running machine learning models. It also calls for an evaluation framework that assesses explanations from both a functional and a human-centered perspective to ensure genuine human understanding of them (Doshi-Velez and Kim, 2017). Responsibilities such as understanding how people interact with these systems, the social aspects of an explanation, legal and ethical implications, and the risks of placing too much trust in AI outputs require deeper engagement than a focus on model accuracy metrics alone. Integrating explainability into the data science curriculum will help ensure that the next generation of data scientists takes responsibility for designing systems that are not only high-performing but transparent and accountable.

3.5.3. Organizational Dimensions

When discussing the role of XAI in promoting social accountability, the technology itself is only part of the broader picture. Equally important is the environment within institutions in which these systems operate. Organizational leadership, internal policies, staff expertise, and incentive structures all influence whether XAI strengthens accountability or merely creates the impression of it (Adadi and Berrada, 2018).

For XAI to serve as a meaningful accountability mechanism, organizations must clearly define who is responsible for decisions made with algorithmic systems. Implementing explainability features alone does not ensure accountability if no one is tasked with reviewing or responding to the explanations that are produced. Organizations should assign individuals or teams clear responsibility for auditing algorithmic outputs and identifying potential problems with them (Doshi-Velez and Kim, 2017). Clear lines of responsibility must also be established so that decision-making authority cannot be easily shifted when issues arise. This is especially important in public-sector organizations, where multiple layers of administration can obscure who ultimately approved a decision, particularly when those decisions impact vulnerable communities. Without clear structures for oversight, even well-designed explainability tools can be overlooked or misused.

Technology alone cannot ensure accountability. It requires people within organizations to understand and interpret algorithmic explanations, and when decision-makers cannot make sense of the information provided by XAI systems, the technology’s practical value is limited. While data scientists and engineers may grasp algorithmic outputs, administrators, managers, and frontline staff often do not. Organizations therefore need training programs that help employees interpret explanations and apply them to actual policy decisions. In addition to technical knowledge, organizations must develop area expertise that connects algorithmic insights with legal requirements, policy goals, and operational realities. Since the gap between technical systems and non-technical decision-makers can be substantial, ongoing professional development is often essential to bridging it (Arrieta et al., 2020). If this gap is not addressed, explanations generated by AI systems may either be accepted without question or dismissed as too complex to understand.

Even when oversight structures and training programs are in place, their effectiveness is limited if the organization resists outside review. In institutions where transparency is seen as a risk rather than a benefit, explainability tools may be reduced to mere compliance requirements rather than serving as meaningful mechanisms for accountability (Mittelstadt et al., 2019). Conversely, organizations that support open evaluation tend to use XAI more productively, applying it to uncover bias patterns, identify unintended consequences, and foster dialogue with communities affected by algorithmic decisions. Rudin (2019) further notes that many organizations rely on black-box machine learning models even when more understandable alternatives are available to them, often reflecting institutional preferences rather than strict technical necessity.

Organizations that regard accountability solely as a regulatory obligation often implement only the minimum transparency required to meet legal standards. Moving beyond minimum compliance requires structures that make accountability costly to ignore. Some organizations establish ethics committees tasked with reviewing algorithmic systems and evaluating their potential risks. Others conduct algorithmic impact assessments both prior to and following use to better understand how these systems affect individuals and communities. Accessible processes for raising concerns are also important, as they enable individuals negatively affected by automated decisions to challenge them (Floridi et al., 2018). Oversight by advocacy organizations and public interest groups can further motivate institutions to take accountability more seriously (Wachter et al., 2017). For AI developers in training, understanding these organizational factors is as important as mastering the technical tools of explainability. A practitioner who can generate a SHAP explanation but does not understand why institutions resist acting on it is only partially prepared for the accountability demands of professional practice.

3.5.4. Political Dimensions

XAI systems do not operate in isolation from politics. Whenever algorithmic tools are used in public programs, such as determining eligibility for government services or informing criminal justice decisions, they operate within broader political environments shaped by power dynamics and competing interests. These political factors affect whether explainability strengthens accountability or merely becomes absorbed into existing institutional frameworks (Diakopoulos, 2016).

A major challenge in algorithmic accountability is that those most affected by automated decisions often have the least ability to challenge them. Automated systems tend to concentrate decision-making authority within institutions, while individuals must navigate complex appeals processes to contest outcomes imposed on them (Eubanks, 2018). In this context, providing explanations alone does not ensure accountability. Even when individuals have the right to receive an explanation, that right is of limited value if they lack the resources or legal support to act on it.

Whether explainability requirements produce meaningful accountability often depends on political will. Laws may reference transparency or the right to explanation, but these principles are sometimes loosely defined or unevenly enforced. Regulations such as the GDPR reference rights related to automated decision-making, yet the practical meaning of these rights remains a matter of ongoing debate (Wachter et al., 2017). Industry influence can also shape regulatory frameworks, sometimes resulting in voluntary guidelines rather than requirements with real consequences for those who ignore them.

Making information about algorithmic systems available to the public does not automatically enable meaningful participation in decisions about them. For XAI to support social accountability, communities affected by algorithmic decisions must have genuine opportunities to question how these systems operate. This includes the ability to examine assumptions within algorithms, challenge classifications applied to them, and contribute to discussions about how automated decisions should be evaluated (Reisman et al., 2018). Meaningful participation is difficult to put into practice, however, because many organizations treat algorithmic outputs as neutral and objective forms of expertise (Pasquale, 2015).

Civil society organizations often play an important role in identifying and challenging harmful algorithmic practices. Advocacy groups, investigative journalists, and independent researchers frequently bring attention to issues that might otherwise remain hidden. Their influence is strongest when independent auditors can examine algorithmic systems without institutional interference and when legal systems allow individuals to challenge automated decisions without excessive cost or complexity (Ananny and Crawford, 2018). Where public oversight is limited, algorithmic systems may reinforce existing power structures rather than challenge them.

These political factors have direct implications for data science education. If future AI developers are trained only to produce technically compliant explanations, meeting regulatory requirements without engaging the communities most affected by algorithmic decisions, then XAI becomes a mechanism for reinforcing existing power structures rather than challenging them. Teaching regulatory awareness must therefore go beyond familiarity with frameworks like the EU AI Act or GDPR. It must include critical examination of how those frameworks were developed, whose interests they protect, and where enforcement falls short of protecting them. Students should understand that the right to an explanation, as established in law, is of limited value to individuals who lack the resources, legal knowledge, and access needed to act on the information they receive (Wachter et al., 2017). They should also understand that civil society organizations have historically played a critical role in bringing attention to algorithmic harms that formal oversight mechanisms failed to catch (Ananny and Crawford, 2018). Preparing AI developers to support, rather than resist, these forms of outside accountability is one of the most significant contributions that socially accountable data science education can make.

3.6. Curriculum Integration Framework

How may the four pillars of socially accountable data science education be actually implemented in an academic course or program? Curriculum integration of this kind requires more than inserting an ethics module into an existing course sequence. As Davis (2020) argues, ethics in data science education must be treated as a substantive component of the development process rather than a standalone lesson appended to technical content. Reviews of XAI pedagogy at other educational levels similarly find that explainability concepts are most effective when embedded directly into subject-matter instruction rather than taught as a separate unit (Prentzas and Binopoulou, 2025). The four pillars described in this paper are designed to be embedded across the data science curriculum – in introductory courses, in capstone experiences, and in the professional norms that programs model through their own documentation and review practices. The following subsections describe how each pillar translates into instructional practice.

The relationship between the four pillars and the four analytical dimensions is not one-to-one: every pillar has a technical, social, organizational, and political expression, and the specific instructional emphasis within each cell varies by course context. Table 1 summarizes this relationship to give instructors a reference for locating, for example, the political dimension of reflexivity as distinct from the political dimension of enforcement.

3.6.1. Answerability

Answerability is operationalized in the curriculum by viewing explanation as a required practice at every stage of development rather than just a bonus feature of a finished AI model. This means that course assignments should require students to produce structured documentation alongside their technical work. Model cards, a tool adapted in the work of Mitchell et al. (2019) and Gebru et al. (2021), provide a ready-made scaffold for this practice: students document intended use cases, known limitations, training data characteristics, and evaluation results in a format designed for non-technical audiences as well as technical ones. Incorporating model card completion as a graded component of project submissions – weighted alongside accuracy metrics rather than treated as supplementary – communicates to students that explanation is part of what it means to build a system responsibly.

Beyond documentation, answerability can be practiced through structured explanation audits in which students are required to apply interpretability tools such as LIME or SHAP to their own models and then present their findings to a peer audience who lack a shared technical background. This exercise develops two competencies simultaneously: the technical skill of generating explanations and the communicative skill of making those explanations meaningful to affected audiences. Chuan et al. (2024) note that XAI approaches themselves can be prone to algorithmic bias, which means explanation audits should also prompt students to interrogate whether the explanations their tools produce are accurate representations of model behavior or potentially misleading ones.

It is important to acknowledge, however, that explanation is not identical to accountability. Producing an explanation, even when that explanation is technically accurate, does not guarantee that the impacted individual can meaningfully understand that explanation, act on it, or use it to challenge a decision. Mittelstadt et al. (2019) caution that explanations in AI systems can create the appearance of transparency while obscuring the deeper structural conditions that produced a decision. An explanation that identifies the features driving a prediction does not necessarily reveal whether those features were appropriate to use, whether the training data was representative, or whether the model’s overall design reflected sound judgment. For this reason, answerability cannot be reduced to the deployment of interpretability tools. It requires sustained engagement with the question of whether explanations are genuinely meaningful to those who receive them. This is a question that connects technical practice to the social, organizational, and political dimensions examined throughout this paper.

3.6.2. Responsibility

Teaching responsibility as a professional habit rather than an abstract ethical concept requires assignments that ask students to reason ahead about harm before a system is deployed rather than retrospectively after consequences have materialized. One practical structure for this is the algorithmic impact assessment, adapted from Reisman et al. (2018), in which students systematically identify who is affected by a system’s outputs, what harms might arise from errors or misuse, and what design choices could mitigate those harms. This assignment works particularly well in project-based courses where students are developing systems over multiple weeks, because it can be revisited and revised as the system evolves, modeling the ongoing nature of responsible practice rather than treating harm assessment as a one-time checklist.

Organizational and societal dimensions of responsibility can be incorporated through case-based instruction. Examining documented cases of algorithmic harm in hiring, healthcare, and criminal justice, among others, gives students a concrete basis for understanding that technical performance and social impact are not the same thing, and that systems can function as designed while still producing unjust outcomes. Elugbaju, Okeke, and Alabi (2024) argue that case-based instructional models support students in developing the structural awareness needed to recognize accountability failures at the level of institutions and systems, not just individual decisions. Pairing case analysis with algorithmic impact assessment exercises connects that structural awareness to the practical choices students are making in their own project work.

3.6.3. Enforcement

The enforcement pillar presents a distinctive pedagogical challenge because it requires translating an institutional concept – accountability with consequences – into a course environment. The key insight here is that course structures themselves can model enforcement mechanisms. When grading rubrics explicitly reward responsible documentation, when projects can be required to undergo revision before approval on ethical grounds, and when peer review is structured around accountability criteria rather than only technical criteria, students experience enforcement as a normal feature of professional practice rather than an external imposition.

Practically, this can be implemented through tiered review processes in which student projects are evaluated not only by the instructor but by structured peer panels that include explicit accountability checkpoints. Before a model can be considered complete, students must demonstrate that they can answer questions about its behavior, its limitations, and its potential for harm – mirroring the kind of institutional oversight that Akinsola (2025) describes as central to governance frameworks for responsible AI. Programs can further reinforce enforcement norms by establishing review mechanisms at the program level, such as an oversight committee that evaluates capstone projects against explicit transparency and fairness standards. Such committees, as Elugbaju, Okeke, and Alabi (2024) suggest, support the development of a structural accountability ecosystem within educational institutions that reflects the governance expectations students will encounter in professional contexts.

Regulatory literacy is also a component of enforcement that the curriculum should address directly. Students need to understand the legal and policy frameworks, including the EU AI Act and the General Data Protection Regulation, that define external accountability requirements for AI systems. Incorporating case-based examination of regulatory compliance failures, and requiring students to evaluate their own projects against relevant regulatory criteria, builds the policy literacy that Chinnaraju (2025) identifies as a core competency for trustworthy AI development.

3.6.4. Reflexivity

Reflexivity is the pillar of this framework that is most easily ignored in a technically oriented curriculum because its outputs are not code or models but habits of critical self-examination. Making reflexivity concrete requires assignments that structure that self-examination and hold students accountable for it. Positionality statements, or brief written reflections in which students identify their own assumptions, values, and potential blind spots before working with a dataset, provide a low-barrier entry point. These statements can be incorporated into project proposals and revisited at project completion, creating a before-and-after record of how students’ understanding of their own influence on the system evolved through the development process.

Reflective design journals, maintained throughout a lengthy project, extend this practice by asking students to document the choices they made, the alternatives they considered, and the reasons they selected one approach over another. This kind of structured reflection develops the reflexivity in computational practice that Cambo and Gergle (2022) describe as essential for developers who need to acknowledge how their modeling choices and personal perspectives shape the systems they build. Stakeholder engagement exercises, in which students present their systems to community members or domain experts who are not data scientists, further develop reflexivity by confronting students with perspectives and concerns that their own positionality may have led them to overlook. Clark et al. (2025) emphasize that meaningful stakeholder engagement requires deliberate structure and preparation, and instructors can provide that structure by scaffolding engagement exercises with clear protocols for listening, documenting feedback, and revising system designs in response.

The stakes of reflexivity failures are not abstract. When developers do not examine their own assumptions, those assumptions are encoded into systems that affect real people. Obermeyer et al. (2019) document a case in which a widely used healthcare algorithm systematically underestimated the medical needs of Black patients because its developers used healthcare spending as a proxy for health need – a choice that appeared technically neutral but reflected unexamined assumptions about the relationship between cost and care. The developers were not acting in bad faith; they were operating within the limits of what they had been trained to see. Reflexivity is the practice of expanding those limits. When developers actively examine their positionality, document their assumptions, and seek perspectives from communities their models will affect, they are more likely to catch the kinds of category errors that produce documented harm. Teaching reflexivity as a core component of data science education is therefore not a soft supplement to technical training. It is a condition of responsible technical practice.

3.6.5. Feasibility of the Proposed Pedagogical Tools

The tools proposed above – graded model cards, structured explanation audits, algorithmic impact assessments, tiered peer-review panels, positionality statements, and reflective design journals – differ considerably in the classroom resources they require and the risks they carry if implemented poorly, and a critical accounting of that variation is necessary before recommending them for adoption.

Model cards and positionality statements are comparatively low-cost: they require no additional software, integrate into existing project deliverables, and can be assessed using a rubric an instructor develops in an afternoon. Their principal risk is superficial compliance, in which students produce a formally complete document without genuine engagement; this is mitigated, though not eliminated, by requiring revision at project completion rather than accepting a single submission at the outset.

Structured explanation audits and algorithmic impact assessments demand more instructional scaffolding. Students need prior exposure to interpretability tools such as LIME and SHAP before they can meaningfully audit their own models, which means these assignments are better suited to courses that already include a technical XAI component than to a standalone ethics module. Impact assessments similarly require some grounding in the domains being examined (e.g., lending, hiring, healthcare) for students to identify plausible harms rather than generic ones; instructors without domain expertise in the case studies they assign may need to draw on outside readings or guest input to keep the exercise substantive rather than formulaic.

Tiered peer-review panels and program-level oversight committees are the most resource-intensive tools proposed here. They require coordination across multiple sections or cohorts, instructor time to train students in evaluating one another’s work against accountability criteria rather than only technical criteria, and, at the program level, institutional buy-in that a single instructor cannot supply alone. Programs without the staffing to support this may still realize much of the pedagogical benefit through a lighter-weight version, such as a single structured peer-review session embedded in an existing capstone course, rather than a standing committee.

Stakeholder engagement exercises carry a distinct feasibility challenge: they depend on the availability of community members or domain experts willing to participate, which is not guaranteed in every institutional setting. Where external stakeholders are not accessible, role-play exercises using documented case studies (e.g., Obermeyer et al., 2019) can approximate some of the perspective-taking benefit, though this is a substitute of lesser fidelity than genuine stakeholder contact.

Taken together, this suggests a staged path for adoption: instructors new to the framework can begin with the lower-cost tools (model cards, positionality statements) before layering in the more resource-intensive ones (peer-review panels, stakeholder engagement) as institutional capacity allows, rather than treating the framework as an all-or-nothing commitment.

4. Discussion

4.1. An Illustrative Example of the Framework in Practice

To illustrate how the four pillars operate together within a cohesive educational and professional context, consider the following scenario. A team of graduate students in a data science program is developing a predictive model intended to flag undergraduate students at risk of academic probation, with the goal of enabling early intervention by advisors. The system is technically functional in that it achieves strong predictive accuracy on a held-out test set. But accuracy alone does not make the system accountable.

Answerability requires that the team be able to explain the model’s outputs to the students it affects and to the advisors who will act on its predictions. In applying SHAP to their model, the students discover that prior GPA and course withdrawal history are the dominant features driving risk scores. Presenting these findings to a peer audience without assuming shared technical background, they realize that withdrawal history may reflect financial hardship or caregiving responsibilities rather than academic disengagement. This is a distinction that the model itself cannot make. Answerability, in this case, emerges a limitation that accuracy metrics obscured.

Responsibility requires the team to reason prospectively about harm before the system is deployed. Completing an algorithmic impact assessment, the students identify that students from lower-income backgrounds are disproportionately represented among those flagged by the model, and that advisor capacity constraints mean that not all flagged students will receive timely outreach. The system could produce the appearance of intervention without its substance, potentially shifting institutional liability while leaving vulnerable students without the support they need. These are important consequences of deployment decisions that responsible developers must anticipate in their systems.

Enforcement requires that these concerns be evaluated by someone beyond the development team. In an educational context, this means a structured peer review process in which classmates evaluate the team’s documentation against explicit accountability criteria — not only whether the model performs well, but whether its limitations are disclosed, its potential for harm is addressed, and its explanations are meaningful to non-technical audiences. The project cannot be considered complete until these questions are answered satisfactorily. This models the institutional oversight structures that AI developers will encounter in professional governance frameworks.

Reflexivity requires thee team to examine the assumptions they brought to the project before it began. In their positionality statements, written at the project’s outset, several students noted that they had not personally experienced academic probation or financial precarity. Revisiting those statements at project completion, they recognize that this positionality led them to frame the problem as a prediction task — one with a clean outcome variable and an obvious technical solution — rather than as a complex institutional and social challenge in which the system’s design choices would themselves shape outcomes. Reflexivity does not invalidate the technical work done by the developers. Rather, it contextualizes this work. In doing so, it makes the team better positioned to communicate its limitations honestly.

Together, these four pillars do not guarantee that the system will be deployed responsibly. But they do ensure that the developers who built it have practiced the habits of thought and documentation that make responsible deployment possible. That should be the goal of socially accountable data science education.

4.2. Applications and Study Limitations

Considered together, the results presented in Section 3 and the scenario in the prior subsection suggest that the four pillars function most effectively not as independent checklist items but as an interlocking set of practices. Answerability without enforcement risks becoming symbolic disclosure; responsibility without reflexivity risks encoding a narrow view of harm; and enforcement without answerability risks becoming compliance for its own sake, decoupled from whether affected communities can actually understand or act on the explanations produced. This interdependence is a central implication of the framework and a design consideration for programs that might be seeking to adopt it. The partial implementation of any single pillar is unlikely to produce the accountability outcomes the framework as a whole is intended to support.

This framework has limitations that should be acknowledged. As a conceptual paper, it has not yet been empirically evaluated in a classroom setting, and the illustrative scenario in Section 4.1, while grounded in documented patterns of algorithmic harm from the literature (e.g., Obermeyer et al., 2019), is a hypothetical composite rather than a reported case. Future work should pilot the curriculum integration strategies proposed in Section 3.6 in actual courses and assess their effects on students’ technical and ethical competencies, using both quantitative measures (e.g., assignment performance, model documentation quality) and qualitative methods (e.g., analysis of student reflexivity statements) to evaluate whether the framework achieves its intended pedagogical outcomes.

4.3. Comparative Positioning within Responsible AI and Ethics Education Scholarship

The framework proposed here is not the first attempt to integrate ethics and accountability content into technical AI education. Prior approaches largely fall into three categories, each of which this framework builds on but also departs from in key ways.

The first category treats ethics as supplementary content: a dedicated module, guest lecture series, or standalone course inserted alongside the technical curriculum (Saltz et al., 2018). This approach has the virtue of being easy to adopt without restructuring existing courses, but it also signals to students, through its placement outside the technical sequence, that accountability is separable from technical competence rather than constitutive of it. The framework proposed in this paper departs from this model explicitly: its four pillars are designed to be embedded within technical coursework (Section 3.6) rather than delivered alongside it.

The second category consists of high-level ethical-principles frameworks, such as the IEEE’s Ethically Aligned Design (IEEE, 2019) or the AI4People framework (Floridi et al., 2018), which articulate values such as fairness, transparency, and beneficence at a level of abstraction intended to apply across institutions and jurisdictions. These frameworks have been influential in shaping policy discourse, but a substantial body of scholarship, including Jobin et al.’s (2019) analysis of global AI ethics guidelines, observes that principles at this level of abstraction rarely specify how a practitioner or an instructor should act differently as a result. The framework proposed here departs from this category by pairing each pillar with specific, assessable classroom instruments (model cards, algorithmic impact assessments, tiered review, positionality statements) rather than stopping at the level of stated values.

The third category is technique-focused XAI pedagogy, which teaches interpretability methods such as LIME and SHAP as technical skills without necessarily connecting them to accountability, governance, or the lived experience of affected communities (see the technique-level literature reviewed in Section 3.5.1). This is the category closest to standard data science curricula, and it is the one most directly extended by this framework: rather than treating interpretability tools as an end in themselves, the answerability pillar (Section 3.1 and Section 3.6.1) explicitly reframes them as instruments in service of a broader accountability relationship, one that this framework argues is incomplete without the responsibility, enforcement, and reflexivity pillars alongside it.

The distinctive contribution of the present framework, then, is not any single pillar or technique in isolation, most of which have some precedent elsewhere, but the explicit structural claim that answerability, responsibility, enforcement, and reflexivity function as an interdependent system (Section 4.2) that must be embedded jointly within technical coursework rather than pursued individually or treated as separable ethical add-ons. This is an empirically testable claim about curriculum design, not merely a restatement of existing values-based frameworks, and Section 4.2 already identifies empirical validation of this interdependence claim as a priority for future work.

4.4. Implementation Challenges and Opportunities

Adopting this framework requires navigating a set of practical obstacles that a purely conceptual treatment can understate. Greatest among these is faculty readiness. Many instructors in data science, computer science, and information science programs were trained primarily in technical methods and may not have formal preparation in accountability theory, XAI critique, or facilitating the kind of reflexive, discussion-based exercises the reflexivity pillar requires (Atenas et al., 2023). Having individual instructors to independently develop this expertise is unrealistic to expect. institutions seeking to adopt the framework should budget for faculty development, such as workshops co-led with colleagues in ethics, science and technology studies, library and information science, or the social sciences, rather than treating implementation as a matter of instructor goodwill alone.

A related risk is resistance from technically oriented faculty who view accountability content as outside their expertise or outside the proper scope of a technical course. This resistance is not necessarily unreasonable, as a faculty member hired and evaluated for technical research productivity has legitimate grounds to question an expectation that they also teach accountability content without additional support, training, or recognition in tenure and promotion processes (Bednarowska-Michaiel & Uprichard, 2026). Programs adopting this framework should treat this as an institutional design problem rather than an individual attitude problem, addressing it through co-teaching arrangements, shared curriculum materials that lower the preparation burden on any single instructor, and explicit recognition of this teaching in workload and evaluation criteria.

Interdisciplinary collaboration, while valuable, is not always institutionally straightforward to develop. Co-teaching across departments can be constrained by budget models that do not easily accommodate split teaching credit, by scheduling conflicts between departments, and by disciplinary differences in how accountability and ethics are conceptualized, which can produce friction even among well-intentioned collaborators. Programs without the administrative flexibility for formal co-teaching can still pursue lighter-weight interdisciplinary input, such as guest sessions, shared case-study materials, or consultation during curriculum design, though this represents a less thorough form of integration than sustained co-teaching (Awuor, 2026).

Resource requirements scale with the tools selected, as noted previously in Section 3.6.5. Programs with limited teaching assistant support or large enrollment sections may find tiered peer-review panels and stakeholder engagement exercises difficult to implement at full scale, particularly in early adoption. In these settings, the staged adoption path suggested in Section 3.6.5, beginning with lower-cost tools before layering in resource-intensive ones, offers a more realistic entry point than attempting full implementation immediately.

None of these challenges is a reason to abandon the framework, but they are important considerations for implementation. A framework that assumes ideal institutional conditions, unlimited faculty preparation time, and frictionless interdisciplinary collaboration offers little practical guidance to programs operating under real constraints. The opportunities implementation offers are correspondingly real: programs that successfully embed even a subset of these practices position their graduates with a form of professional preparation, demonstrable accountability literacy, that is increasingly expected by employers and regulators alike, and that few competing programs currently provide in a structured way.

5. Conclusions

This paper has proposed a framework for integrating explainable AI (XAI) and social accountability into data science education, structured around four interconnected pillars – answerability, responsibility, enforcement, and reflexivity – and situated within technical, social, organizational, and political dimensions of XAI implementation. Taken together, these elements constitute a pedagogical approach that moves beyond narrow technical training to cultivate practitioners who understand and accept their role as stewards of consequential sociotechnical systems.

The central argument of this framework is that explainability, and AI ethics at large, is not a feature to be appended to a finished model but a professional obligation to be embedded throughout the development lifecycle (Davis, 2020). Teaching students to document their data choices, interrogate the assumptions embedded in their models, and communicate uncertainty and limitations to diverse audiences is not a diversion from technical education. Rather, it is a deepening of it. A data scientist who cannot explain a model’s outputs to those affected by its decisions, or who has not considered who might be harmed by a system’s errors, is not fully prepared for professional practice.

The development of AI systems that are trustworthy, fair, and responsive to the communities they serve is one of the defining challenges of the generative AI era. That challenge will be met, or not met, primarily through the choices made by the practitioners who are being trained today. Data science education has an important role to play in shaping those choices. By treating social accountability not as a constraint on technical innovation but as one of its highest aspirations, programs can help ensure that AI development serves not just the interests of those who build these systems, but the interests of all who are shaped by them.