Refine: use inherent LLM limitations as organic adversarial examples for critique exercises¶

Strategy · All strategies
Evidence · no claims cited

Description¶

The fourth framework component leverages the inherent limitations of current LLMs — cultural biases, hallucinations in complex medical cases, and confident but incorrect citations — as organic adversarial examples for students to critique, rather than artificially engineering misinformation. The authors argue that 'finding and refuting AI hallucinations turns the tool’s weakness into a critical thinking exercise, transforming the interaction from passive consumption to active evaluation'.

Design Implications¶

Context¶

Requirements¶

• Access to AI-generated explanations containing authentic limitations such as hallucinations or confident but incorrect citations that students can be asked to identify and refute.

Constraints¶

Target Learners¶

• medical students
• students evaluating AI-generated explanations

Target Learning Goals¶

• critical thinking
• active evaluation of AI output
• verification against sources

Related Strategies¶

• Realign: deploy Socratic withholding protocols that delay direct answers for conceptual or diagnostic queries
• Require: process-based assessment through metacognitive rationales for accepting or revising AI suggestions

Examples¶

-

Key Sources¶

• Wang H and Shan W (2026). The safety gap: restoring productive struggle through pedagogically aligned generative AI. Front. Educ. 11:1757622. https://doi.org/10.3389/feduc.2026.1757622