Study Finds AI Explanations Raise False Negatives in 228-Person Innovation Test
Updated
Updated · Computerworld · Aug 20
Study Finds AI Explanations Raise False Negatives in 228-Person Innovation Test
3 articles · Updated · Computerworld · Aug 20
Summary
228 experienced evaluators reviewing nearly 50 MIT challenge submissions were more likely to make worse calls when an LLM explained its pass-fail recommendation, especially by rejecting ideas human experts would have advanced.
67% of LLM recommendations were accepted overall; reviewers agreed with AI about 75% of the time but matched human-only decisions just 54%, showing how strongly the tools swayed judgment.
Narrative rationales were the main problem: they suppressed productive overrides, gave reviewers ready-made reasons to reject proposals, and amplified negativity bias, while black-box recommendations improved alignment with expert baselines.
The researchers from Harvard, MIT and the University of Washington said enterprises should treat AI explanations as behavioral interventions, using testing, uncertainty disclosures or even simpler recommendations in high-stakes screening tasks.