OpenAI Research: Why Language Models Hallucinate

Hallucinations are driven by evaluation incentives

Language models hallucinate because current evaluation methods reward guessing over honesty about uncertainty. When models are graded primarily on accuracy—the percentage of questions answered correctly—they are incentivized to guess even when they lack the necessary information, as a correct guess improves the score while admitting uncertainty (abstaining) guarantees zero points.

OpenAI's research indicates that errors are more detrimental than abstentions. According to the OpenAI Model Spec, it is preferable for a model to indicate uncertainty or request clarification than to provide confident but incorrect information. However, because accuracy-only scoreboards dominate model cards and leaderboards, developers are motivated to build models that guess rather than hold back.

Comparing Accuracy vs. Hallucination Rates

A comparison between gpt-5-thinking-mini and OpenAI o4-mini demonstrates the trade-off between strategic guessing and reliability:

Metric	gpt-5-thinking-mini	OpenAI o4-mini
Abstention rate (no answer given)	52%	1%
Accuracy rate (right answer)	22%	24%
Error rate (wrong answer)	26%	75%

While OpenAI o4-mini has a slightly higher accuracy rate, its error rate (hallucination rate) is significantly higher because it rarely abstains.

The origin of hallucinations in next-word prediction

Factual inaccuracies originate from the statistical nature of pretraining. During pretraining, models learn by predicting the next word in vast amounts of text without "true/false" labels. While consistent patterns—such as spelling and parentheses—are easily learned and errors in these areas disappear with scale, arbitrary low-frequency facts (e.g., a specific person's birthday) cannot be predicted from patterns alone.

Because models only see positive examples of fluent language and must approximate the overall distribution, they struggle to distinguish valid statements from invalid ones without examples labeled as invalid. This statistical mechanism leads to the hallucinations that occur when a model attempts to predict a low-frequency fact it cannot reliably determine.

Correcting the evaluation framework

To reduce hallucinations, OpenAI proposes updating widely used accuracy-based evaluations to discourage guessing. This involves:

Penalizing confident errors more heavily than uncertainty.
Providing partial credit for appropriate expressions of uncertainty.

OpenAI asserts that adding a few uncertainty-aware tests is insufficient; the primary scoreboards must be reworked to reward humility and discourage blind guessing to broaden the adoption of hallucination-reduction techniques.

Addressing common misconceptions about hallucinations

OpenAI's research clarifies several key points regarding the nature of AI hallucinations:

On Accuracy: Accuracy will never reach 100% because some real-world questions are inherently unanswerable regardless of model size or reasoning capabilities.
On Inevitability: Hallucinations are not inevitable because models can be designed to abstain when uncertain.
On Model Size: Avoiding hallucinations does not exclusively require the intelligence of larger models; it can be easier for a small model to know its limits (e.g., a small model with no knowledge of a specific language can simply abstain).
On the Nature of the Problem: Hallucinations are not "mysterious glitches" but result from understood statistical mechanisms and reward structures in evaluations.
On Measurement: Simply creating a "hallucination eval" is not enough; all primary evaluation metrics must be updated to reward expressions of uncertainty.