OpenAI researchers compared large language models to students taking a test. When LLMs face a multiple-choice question they cannot answer, they often guess instead of admitting uncertainty. This behavior mirrors how students might guess to avoid leaving a question blank.

The paper explains this as a result of binary evaluation metrics. Models earn a point only for correct answers, making a wrong guess mathematically better than stating they don't know. This creates what researchers call an 'uncertainty penalty epidemic,' systematically incentivizing hallucinated responses.

This insight reframes how developers might approach AI training and evaluation. Moving beyond simple right/wrong scoring could encourage models to express confidence levels or abstain from answering. The discussion raises questions about designing better benchmarks for reliable AI systems.