Why Language Models "Hallucinate" Answers – And Why It Happens At All

Why Language Models "Hallucinate" Answers

OpenAI published research in September 2025 explaining that hallucinations happen not because models are broken, but because the way they are trained and evaluated incentivizes guessing over admitting uncertainty.

What Are "Hallucinations"?

A hallucination is when a language model states something that sounds convincing but is simply wrong, sometimes even on basic questions.

The Root Cause: Training Incentives

Many benchmarks reward a "hit" on an answer, even a guess, more than an "I don't know." Like a multiple-choice test: guessing gives you a 1-in-365 chance of being right, while abstaining guarantees zero points.

The Fix

OpenAI proposes:

• Penalizing confident wrong answers more heavily
• Giving partial credit for expressing uncertainty instead of guessing

The key is updating the central benchmarks everyone uses, as long as leaderboards reward guessing, models will keep learning to guess.