Why Language Models Hallucinate
A research paper published by OpenAI on the causes of hallucination in large language models, identifying as a systemic source the fact that scoring mechanisms reward guessing rather than acknowledging not knowing.
Timeline (last 90 days)
The paper gives a lower bound: after pretraining, the hallucination rate is at least equal to the proportion of facts in the training set that appear only once.
The paper's core argument is that as long as the main leaderboards look only at accuracy, guessing remains a rational strategy for models.
Sources cited in the article include Kalai et al., "Why Language Models Hallucinate" (arXiv:2509.04664), an official OpenAI article, Ji et al.'s hallucination survey (ACM Computing Surveys), and the SimpleQA data in the GPT-5 system card.
OpenAI explained in a September 2025 paper, "Why Language Models Hallucinate", that most training and benchmarks only score answers as right or wrong and give no credit for "I don't know", which leads models to lean toward guessing.
All Facts (4)
The paper gives a lower bound: after pretraining, the hallucination rate is at least equal to the proportion of facts in the training set that appear only once.
50%UnverifiedThe paper's core argument is that as long as the main leaderboards look only at accuracy, guessing remains a rational strategy for models.
50%UnverifiedSources cited in the article include Kalai et al., "Why Language Models Hallucinate" (arXiv:2509.04664), an official OpenAI article, Ji et al.'s hallucination survey (ACM Computing Surveys), and the SimpleQA data in the GPT-5 system card.
50%UnverifiedOpenAI explained in a September 2025 paper, "Why Language Models Hallucinate", that most training and benchmarks only score answers as right or wrong and give no credit for "I don't know", which leads models to lean toward guessing.
50%