A brief reflection reduced acceptance of incorrect AI advice
Does asking students to pause and examine AI advice help them resist plausible mistakes? A randomized experiment with 342 undergraduates compared independent work, open ChatGPT support and ChatGPT with a reflection prompt. Students completed reasoning and information-evaluation tasks, recording initial and final judgments. Those with open support accepted incorrect recommendations on 62.4% of relevant trials; reflection reduced that figure to 39.7%. The groups encountered similar rates of correct AI advice, making unequal exposure an unlikely explanation. Reflection also improved awareness of reliance. The result supports a deliberate checking step during assisted work, while leaving unanswered whether the benefit persists in later, unaided learning.
Why it matters for leaders
Pilot a reasoning checkpoint in an existing lesson: students record an initial answer, identify evidence for accepting or rejecting advice, and explain their final choice. Compare that work with a later task completed without AI. Track reasoning quality and teacher review time alongside answer accuracy.
The evidence
The incorrect-advice comparison yielded an odds ratio of 0.40, with a 95% confidence interval of 0.28-0.56. Chat logs supplied behavioral measures.
Who was studied
Undergraduates; indirect evidence for K-12.
What the evidence cannot tell us
One Chinese university; immediate tasks, not long-term learning. Calibration partly overlaps reliance. Exploratory mediation does not establish causation.
Read the original sources
Summary and leadership analysis from the Week ending 2026-09-18 brief. The edition’s selection notes explain source access and evidence limitations.