Evaluating large language models for accuracy incentivizes hallucinations

A study published in Nature on 22 April 2026 reveals that evaluating large language models for accuracy can lead to an unintended increase in hallucinations, where models generate false information to appear accurate.