Audit educational LLM deployments for differential treatment before serving marginalized learners¶

Strategy · All strategies
Evidence · no claims cited

Description¶

The authors recommend that "Educational deployments require urgent auditing, as safety alignment may actively harm vulnerable populations." Concretely, they propose "developing culturally-equitable alignment to maintain educational depth across all socio-economic profiles," and call for future research extending audits to other cultural contexts and to multi-turn conversational tutoring. The strategy treats persona-based API auditing as a pre-deployment safeguard for conversational tutors.

Design Implications¶

Context¶

Requirements¶

• Controlled API access to deployed models; persona-based test protocols with metrics for refusal, complexity, framing, and epistemic validation

Constraints¶

• The audit's proof-of-concept covered a single event and single-turn interactions; multi-turn drift remains untested

Target Learners¶

• Marginalized students, including ethnic minorities and low-resource vocational learners

Target Learning Goals¶

• Equitable access to full historical complexity and civic interpretive tools from AI tutors

Related Strategies¶

• Mitigating Racial Bias in Edtech Products
• AI in Education Toolkit for Racial Equity
• Evaluate AI systems in DLPs for fairness across learner groups, not just overall accuracy
• Four-step deployment preparation for patient-facing generative AI: audit, triage, fill priority gaps, and retain evidence

Examples¶

-

Key Sources¶

• Alexis Popovici, Andrei Ionascu, Adrian Marius Dumitran. (2026). The Paternalistic Filter: Epistemic Injustice and Differential Refusal in LLM-Mediated History Education for Marginalized Romanian Students. arXiv preprint. https://arxiv.org/abs/2607.11292