Why AI Language Models Hallucinate: The Research Behind False Answers

An AI hallucination is a statement produced by a language model that sounds fluent and confident but is factually wrong — a fabricated citation, a wrong date, a person who does not exist. In a research paper published in September 2025, OpenAI researchers describe hallucinations in these terms and state that, with current training and evaluation methods, the problem cannot be fully solved.

How the training process creates errors

According to the OpenAI paper, posted on arXiv, part of the explanation lies in how models are pretrained. During pretraining, a model learns only from fluent text, without labels marking which statements are true or false. It is trained with a statistical objective (cross-entropy loss) that rewards predicting likely next words, not verifying facts.

The researchers show this produces errors even on data that itself contains no mistakes. For facts that appear only once in the training data — the researchers give the example of a person's birthday — the paper's analysis finds that the hallucination rate on those facts should be at least as high as the share of such "singleton" facts in the training set. In other words, if a fact was mentioned only once, the model is mathematically prone to getting it wrong. The paper calls some errors unavoidable even for models with no limits on computing power, because certain tasks are what it describes as computationally intractable.

Why grading systems reward guessing

The paper's second argument concerns how models are evaluated after pretraining. It compares a language model to a student facing a multiple-choice exam: a blank answer scores zero, so guessing is always the statistically better strategy when every wrong answer costs the same as leaving the question unanswered. OpenAI's researchers examined ten widely used AI benchmarks — including GPQA, MMLU-Pro and SWE-bench — and found that nine of them use binary grading that gives zero credit for an answer like "I don't know," the same score as a wrong answer.

That scoring pattern, the paper argues, pushes models to keep guessing rather than admit uncertainty, because confident guessing scores better on average than honest hedging. TechCrunch, reporting on the paper, described this as evaluations that create the wrong incentives by rewarding lucky correct guesses the same way as genuinely well-supported answers.

Reasoning models have hallucinated more, not less

The trend is not limited to older models. OpenAI's own system card for its o3 and o4-mini reasoning models, released in April 2025 and reported by TechCrunch, found that hallucination rates rose with newer reasoning systems on the company's internal PersonQA benchmark, which tests factual questions about people.

OpenAI said more research is needed to understand why hallucinations increase as reasoning models scale up, noting that these models make more claims overall, which produces both more correct and more incorrect statements. Neil Chowdhury, a researcher at the AI research group Transluce quoted by TechCrunch, offered one hypothesis: the type of reinforcement learning used to train o-series reasoning models may amplify issues that standard training techniques normally mitigate.

What OpenAI proposes to fix it

The September 2025 paper proposes changing how benchmarks are scored rather than adding separate hallucination tests. Its suggested fix is to build explicit confidence thresholds into grading instructions — for example, telling a model to answer a question only if it is more than 75% confident, with mistakes costing more points than an abstention. In theory, this makes abstaining the mathematically better choice when a model is unsure, rather than the worse one, as happens in current scoring systems.

The paper is explicit that this approach would reduce, not eliminate, hallucinations. Because some errors stem from statistically rare facts or from tasks the researchers describe as computationally intractable, the authors conclude that a language model with zero hallucinations is not achievable with current methods.