Why does AI hallucinate, and what actually reduces it?

Where hallucinations come from

Pretraining teaches a model to predict the next token across a large corpus of fluent text. The Why Language Models Hallucinate paper frames the consequence plainly: the model sees only positive examples of language and has to approximate the whole distribution, with no true or false labels attached to any statement. The authors argue that hallucinations then originate as ordinary errors in binary classification, and that if incorrect statements cannot be distinguished from facts, they arise from natural statistical pressure rather than from a mysterious defect.

The paper uses a useful analogy for why some facts are harder than others. Label millions of pet photos as cat or dog and a classifier learns the distinction reliably. Label each photo with the pet's birthday and errors are guaranteed no matter how good the algorithm is, because the target is arbitrary relative to the input. Rare, arbitrary facts — a specific person's dissertation title, a specific case citation — sit in that second category.

That is the analysis part. The reported behaviour is concrete: OpenAI's accompanying post describes asking a widely used chatbot for the dissertation title of one of the paper's own authors and getting three different answers, none correct, and three different wrong birthdays.

Why the problem persists even as models improve

The paper's second argument is about grading, not architecture. Most benchmarks report accuracy: the share of questions answered exactly right. Under that scoring, guessing has positive expected value and abstaining has none. A model that says "I do not know" when it does not know scores worse on the leaderboard than one that guesses, so training pressure favours the guesser.

The SimpleQA figures published in the OpenAI post are the clearest illustration available. gpt-5-thinking-mini abstained on 52 percent of items, was accurate on 22 percent, and was wrong on 26 percent. OpenAI o4-mini abstained on 1 percent, was accurate on 24 percent, and was wrong on 75 percent. The older model wins on accuracy by two points and loses on error rate by 49.

The authors' proposed fix is socio-technical rather than architectural: change the scoring on the mainstream evaluations that dominate leaderboards, penalising confident errors more than expressions of uncertainty, rather than adding another standalone hallucination benchmark that hundreds of accuracy-only evals will drown out. Whether the wider field adopts that is an open question, not a settled outcome.

The mitigations with real mechanisms behind them

Retrieval grounding is the first lever. Google's Gemini API documents a Grounding with Google Search tool whose stated purpose is to reduce model hallucinations by basing responses on real-world information and to return citations for the claims. Grounding narrows the space of things the model has to recall from weights, but it moves the failure mode rather than deleting it: a bad retrieval produces a confidently wrong grounded answer.

Citation binding is the second. Anthropic's Citations feature chunks supplied documents into sentences and returns responses where each claim carries citations pointing at specific character ranges, page numbers, or content-block indices in the source. The practical value is auditability — a reader or a downstream check can verify each sentence against the span it claims to come from, instead of trusting the paragraph as a whole.

Automated grounding scoring is the third and least used. Google Cloud's check grounding API takes an answer candidate plus up to 200 supplied facts and returns a support score from 0 to 1 indicating how much the candidate is entailed by those facts, plus per-claim citations. Its documented rule is strict: a claim that is only partially entailed counts as ungrounded, and the example given is "Google was founded by Larry Page and Sergey Brin in 1975" — correct founders, wrong year, therefore ungrounded as a whole. Google documents latency under 500 ms and describes setting a citation threshold to filter likely-hallucinated claims at inference time.

What this looks like in a real workflow

Put the mitigations in order of cost. Start by supplying the source material instead of relying on model memory, because recall of rare facts is exactly the weak case the research describes. Then require citations at the claim level, not the answer level. Then add a programmatic grounding or entailment check if wrong answers are expensive, and treat a low support score as a routing signal to a human rather than as a number to display.

Prompting the model to say when it is unsure helps at the margin and is worth doing, but it is working against the training incentive the paper describes rather than with it. Do not treat it as a control.

Keep verification proportional to consequence. Free-text brainstorming needs none of this. Anything that will be filed, published, prescribed, or paid against needs a human who checked the cited source, because the cheap mitigations all fail quietly rather than loudly.

The cost of skipping verification

The legal profession provides the clearest public record of what unverified output costs, because court decisions are published. Damien Charlotin's AI Hallucination Cases database tracks decisions in which a court explicitly found or implied that a party relied on hallucinated material — typically fabricated citations. As of its 14 September 2026 update the database listed 2,041 identified cases, with the largest counts in the USA, Canada, Australia, UK and Israel, and with pro se litigants and lawyers accounting for most of the entries.

Read that as a record of a verification failure mode rather than as a measure of how often models hallucinate. It counts only cases that reached a decision and where the court addressed the AI use directly; the database's own description says it does not track the wider universe of fake citations. It is licensed CC BY 4.0 and published with a downloadable CSV, so the figure is checkable.

Limits of this answer

RepoRadar has not independently benchmarked hallucination rates for any model, and this article does not rank models on truthfulness. The SimpleQA numbers quoted here are OpenAI's own published figures for two of its models on one evaluation, not a cross-vendor comparison.

Vendor grounding and citation features are described from their current documentation, not from hands-on measurement of how well they work on your data. Their stated purpose is reduction, not elimination, and none of the three vendors claims otherwise.

The research argument about evaluation incentives is a well-argued position from named researchers, not a settled consensus of the field. Treat it as the best current explanation rather than as a proven law.

A useful next action

Take one real task where a wrong answer would cost you something, and run it twice: once with the model answering from memory, once with the source documents supplied and citations enabled. Then check every cited claim against the span it points at. The gap between the two runs, measured on your own material, is a better guide to how much grounding buys you than any published rate.