Independent briefing Reliability & safety
Why language models hallucinate
A training-and-evaluation explanation for why models often guess instead of admitting uncertainty.
Independent briefing of OpenAI (2025). Read the original for its full evidence, context, and wording.
Start here
Three questions this article answers
Why do fluent models produce false statements?
Next-token training rewards plausible continuation, while many evaluations reward an attempted answer more than an honest “I don’t know.”
Are hallucinations only a data-quality problem?
No. Even with better data, uncertainty and evaluation incentives influence whether a model guesses.
What would improve model behavior?
Evaluations and product incentives that reward calibrated uncertainty, verification, and appropriate abstention.
Go deeper
Three related questions it doesn’t answer—matched to established papers
Can uncertainty detect confabulation without knowing the correct answer?
The method measures disagreement in meaning across sampled answers.
Paper answering this questionDetecting Hallucinations in Large Language Models Using Semantic EntropyDo people interpret model confidence accurately?
Experiments examine calibration and how users perceive expressed confidence.
Paper answering this questionWhat Large Language Models Know and What People Think They KnowCan language-model confidence be calibrated?
The study compares confidence signals and calibration methods across question-answering tasks.
Paper answering this questionHow Can We Know When Language Models Know?The useful summary
What to take away
The article explains hallucination as more than a mysterious defect. A language model learns to continue patterns, not to maintain an internal fact ledger. Later evaluations often turn uncertainty into a bad strategy by scoring a guess as potentially correct and an abstention as automatically wrong.
That diagnosis points to practical changes. Systems should separate generation from verification, show sources where possible, and be allowed to say that evidence is insufficient. Benchmarks should use selective accuracy and calibration—not only the number of questions answered—to reward dependable behavior.
Final conclusion
The three answers, brought together
Models hallucinate because plausible generation and answer-at-all-costs evaluation favor guessing. The problem is not solved by cleaner data alone. Better incentives must reward uncertainty, checking, and refusal when the evidence is weak.