Toward socially accountable data science education: A conceptual framework for integrating explainable AI and accountability principles

Synthesis: Lund and colleagues (2026) argue that programs training AI developers teach technical proficiency while leaving social accountability to elective ethics content, and propose a conceptual framework for embedding both in the core curriculum of CS education. Its structure is four pillars (answerability, responsibility, enforcement, and reflexivity), each crossed with four dimensions of XAI implementation: technical, social, organizational, and political. The distinctive move is pedagogical concreteness: graded model cards, explanation audits with LIME and SHAP, algorithmic impact assessments, tiered peer-review panels, positionality statements, and reflective design journals, with a staged path for programs that cannot afford the full set. The authors are explicit that the framework is conceptual and untested, and that explanation alone is not accountability: an accurate description of a model's behavior says nothing about whether the right features should have been used.

Key Findings

1. Accountability is taught as optional content. Global AI ethics frameworks name transparency, fairness, and accountability almost universally, but the usual curricular response, a guest lecture or a standalone course, signals that accountability is separable from technical competence.
2. Four pillars are crossed with four dimensions. Answerability, responsibility, enforcement, and reflexivity each take a technical, social, organizational, and political expression,
3. Enforcement is the layer education most often omits. Grading rarely rewards documentation, so the framework proposes rubrics, projects returned on ethical grounds, accountability peer review, and regulatory literacy in the EU AI Act and the GDPR.
4. Reflexivity is dropped first because it yields habits, not artifacts. Positionality statements and design journals make it gradeable, and the authors call it unrealistic to expect instructors to build these exercises alone without faculty development.
5. The assignments are rated for classroom cost and failure mode. Model cards and positionality statements fit work instructors already assign but risk superficial compliance, while audits, impact assessments, and oversight panels need Scaffolding or budget.
6. Explanation alone is not accountability. An accurate description of a model's behavior says nothing about whether the right features should have been used, and the framework is an untested design proposal rather than a validated instrument.

The gap, and the four pillars

Global AI ethics frameworks assert transparency, fairness, and accountability almost universally while supplying few enforcement mechanisms, and the curricular response treats accountability as separable from competence. The framework's answer is four pillars. Answerability obliges developers to explain what data a system relies on and how it produces outputs. Responsibility is a lifecycle commitment shared unequally across developers, organizations, and governments, illustrated by the medical risk model that used healthcare spending as a proxy for need and underestimated Black patients' needs, bias reaching a system through a design choice rather than a bug. Enforcement supplies consequences: rubrics, projects returned on ethical grounds, peer review, and regulatory literacy. Reflexivity yields habits rather than artifacts, through positionality statements and design journals. The pillars work as an interlocking set: answerability without enforcement risks symbolic disclosure.

Four dimensions, and what instructors assign

Each pillar has a technical, social, organizational, and political expression, so answerability runs from interpretability outputs (LIME, SHAP, counterfactuals) to explanations legible to non-technical stakeholders, with regulatory literacy alongside technical documentation. The framework's classroom forms are concrete:

• Graded model cards document intended use, limitations, training data, and evaluation results for non-technical readers, weighted alongside accuracy.
• Explanation audits have students apply LIME or SHAP and present results to a non-technical audience, asking whether an explanation is faithful or merely persuasive.
• Algorithmic impact assessments ask who is affected, what harms misuse could cause, and which design choices would mitigate them, revised as a project evolves.
• Tiered peer review, positionality statements, and design journals grade behavior and potential for harm, and surface concerns a student's positionality hid.

Feasibility, and an illustrative scenario

Model cards and positionality statements are low-cost and integrate into existing deliverables, at the risk of superficial compliance. Explanation audits and impact assessments need scaffolding: prior exposure to interpretability tools and enough domain grounding (lending, hiring, healthcare) for harms to be plausible. Tiered panels and oversight committees cost the most, though one peer-review session inside a capstone captures much of the benefit. In the illustrative scenario, a graduate team predicts which undergraduates risk academic probation; SHAP surfaces prior GPA and withdrawal history, a non-technical audience sees that withdrawal may reflect financial hardship or caregiving, an impact assessment finds lower-income students over-represented among those flagged, and positionality statements reveal that several had framed the problem as prediction rather than an institutional challenge.

How the framework was built

This is a narrative synthesis rather than a systematic review: an interpretive, argument-building approach instead of reproducible retrieval, with no fixed date range, inclusion criteria, or dual-reviewer screening. It draws on accountability theory (Bovens' distinction between explanation and consequence), the XAI literature questioning whether explanation guarantees accountability, and scholarship on ethics curricula. It positions itself against ethics as supplementary content and against principle frameworks such as the IEEE's Ethically Aligned Design.

What this means for practice

• Instructors. Add graded model cards and positionality statements to deliverables you already assign, require a final revision rather than a single submission, and send projects back on ethical grounds.
• Administrators. Sequence the rest outward: explanation audits and impact assessments only after exposure to LIME or SHAP and enough domain grounding for harms to be plausible, and budget for faculty development and co-teaching.
• Researchers. Treat it as a design proposal to pilot rather than evidence; nothing has been tested with students, and the authors call for evaluation using assignment performance, documentation quality, and reflexivity statements.

Limitations

• Not evaluated in a classroom: the authors present a conceptual paper and call for piloting these strategies in real courses.
• The discussion scenario is a hypothetical composite built from documented patterns of harm, not a reported case.
• The synthesis is narrative, not systematic, so coverage is argument-driven rather than reproducible.

Citation

Lund, B. D., Alizai, K., Amoje, E., Chandrasekaran, A., Darvischi, S., Denmark, J., Paulraj, N. J., Nallamothula, L., Paes, A., & Vemulapalli, B. S. (2026). Toward socially accountable data science education: A conceptual framework for integrating explainable AI and accountability principles. AI in Education, 2(3), 32.