Why language models hallucinate

Source: OpenAI Date: 2025-09-05 URL: https://openai.com/index/why-language-models-hallucinate

Summary

Summary

OpenAI published an accessible explanation of why language models hallucinate — describing the mechanistic reasons that models confidently generate false information, covering the training process, the architecture’s relationship to uncertainty, and the conditions under which hallucinations are most likely to occur. The piece was aimed at a general technical audience rather than researchers.

Implications

Research/safety thread. Publishing “why models hallucinate” is an educational and trust-building move: it signals that OpenAI understands the problem mechanistically, not just as a product defect to be patched. The audience is developers and enterprise buyers who need to understand and design around hallucination risk. The explanatory framing — this is why it happens, not just that it happens — is more useful for building reliable applications than product disclaimers. The September 2025 timing coincides with DeepResearch and other high-stakes information tasks becoming common, where hallucination costs are particularly high. The implication is that OpenAI was managing user expectations as much as explaining the technical reality: hallucination is a feature of the architecture, not a bug being fixed, and deployers need to account for it.