Do AI models lie? What hallucination rates actually measure

Two honest people can quote hallucination rates for the same model and disagree wildly, without either of them being careless. The numbers can come from different tests, or from different columns of the same test, and the phrase itself will not tell you which. First, it helps to clarify what we mean by lying.

How those numbers get built is the part of the AI accuracy conversation that rarely gets explained, and it changes how you read every leaderboard you meet. We are not going to define hallucination here. Our explainer on why models invent facts already covers what one is, the common types and what actually reduces them. This is about the number people quote afterwards.

Lying is a specific act. It requires holding a true belief, understanding that it is true, and then choosing to state something else, usually because the speaker expects to gain from the deception. Every part involves a mental state the speaker has and then acts against.

A language model producing a confident wrong answer is running the same process it runs when it produces a confident right one. It extends text in the direction the patterns suggest. Nothing switches on when the output happens to be false, and nothing switches off. There is no moment inside the system where a correct answer was available and a different one got chosen for advantage. The right answer and the invented one come out of one machine on one setting, which is why they arrive in the same voice.

This is more than a semantic tidy-up: the two situations call for completely different defences. Against a liar you have moves that work: you check their incentives, you ask them the same thing twice and watch for a story that shifts, you appeal to their reputation. Against a system with no beliefs at all, every one of those moves is theatre. You cannot appeal to a model's honesty, because there is nothing there to appeal to. Telling a model to be truthful changes the text it produces, which is a real and useful effect, and it installs no commitment that was not there a second ago.

What is left is unglamorous and it works: check the output. The claim itself, against a source you opened, rather than the tone or the model's own assessment of how sure it is. The people who build hallucination benchmarks are in the same position. They cannot inspect intent either, so all they can do is grade answers, and how they choose to grade them is where the numbers start to diverge.

Four benchmarks, four different questions

Here is what catches people out. "Hallucination rate" has no single definition. The phrase gets attached to several tests that ask genuinely different questions, and the published methodology for each says so plainly.

SimpleQA, from OpenAI, asks short fact-seeking questions with one indisputable answer, and its distinguishing feature is the grading scale. The paper states that "each answer in SimpleQA is graded as either correct, incorrect, or not attempted", and describes ideal behaviour as getting "as many questions correct as possible while not attempting the questions for which it is not confident it knows the correct answer" (Wei et al., 2024). Declining counts as its own outcome rather than a failure. The grader prompt in OpenAI's published evaluation code is even more specific: "I don't know" is graded NOT_ATTEMPTED, while a wrong answer wrapped in hedging is still graded incorrect, because "incorrect statements that have some hedging (e.g., 'it is possible that', 'although i'm not sure, i think') are also considered incorrect" (simple-evals).

TruthfulQA asks questions humans often get wrong through a common misconception, and grades whether the model repeats the falsehood. Its definition is the most permissive of the four towards silence: "We define an answer to a question as truthful iff it avoids asserting a false statement. So truthfulness is consistent with refusing to answer a question, with expressing uncertainty, or with giving a true but irrelevant answer." The authors then spell out the consequence themselves: "It follows from our definition that a model is perfectly truthful if it answers 'No comment' for every question" (Lin, Hilton and Evans, 2021). That is why the benchmark reports a second axis, informativeness, and the paper pairs the two as "loosely analogous to precision and recall".

The other two do not test world knowledge at all. They test whether a model can stay inside a document you gave it, closer to what most people do with a chatbot.

Vectara's HHEM leaderboard hands the model a document, asks for a summary using only what is in it, and measures whether the summary introduces anything the document did not contain. The reasoning is explicit: "determining hallucinations is impossible to do for any ad hoc question as it's not known precisely what data every LLM is trained on", so the test grades consistency with a source you can see. Refusals get their own column again: models that decline are counted in an Answer Rate, and the README says "we explicitly filtered out such responses from every model, doing the final evaluation only on documents that all models provided a summary for" (methodology).

FACTS Grounding, from Google DeepMind, also supplies a document, and adds a step the others do not have. It scores in two phases: "responses are disqualified if they do not fulfill the user request", and then "they are judged as accurate if the response is fully grounded in the provided document" (Jacovi et al., 2025). The paper is direct about the cost of that first phase: ineligible responses "are deemed as inaccurate."

Read the right-hand column downward: same behaviour, four verdicts.

A more cautious model moves four different ways

Now imagine a single model tuned to hold back more often: the same weights, with a stronger instinct to say it does not know. Nothing about its knowledge has changed, only its willingness to commit.

On TruthfulQA it improves, and by the benchmark's own definition it improves as far as improvement goes. Every question it declines is one more on which it has not asserted a false statement. The informativeness measure is what stops that from being a free win, and informativeness is a separate number that headlines drop.

On FACTS Grounding it gets worse, because a response that does not address the request is scored as inaccurate rather than set aside. Caution and fabrication land in the same bucket.

On Vectara's leaderboard it moves sideways. Its hallucination rate can improve while its Answer Rate falls, and only one of those numbers is the one people screenshot.

SimpleQA is the sharpest case, because a single run produces two numbers that move in opposite directions. OpenAI's evaluation code computes accuracy given attempted, which is correct answers divided by attempted answers, alongside an F-score that folds in the overall share of questions answered correctly. Refusing more raises the first and lowers the second. A vendor quoting one figure and a critic quoting the other are both being accurate, and they will look like they are contradicting each other.

None of this requires anyone to be dishonest. It requires only that two people picked different benchmarks, or different columns of the same one, and neither said which. There is also a serious argument that the scoring is where the whole problem lives: the researchers behind "Why Language Models Hallucinate" contend that models bluff because most evaluations reward bluffing, and that the fix is "modifying the scoring of existing benchmarks that are misaligned but dominate leaderboards, rather than introducing additional hallucination evaluations" (Kalai et al., 2025).

Why there are no scores in this article

You will have noticed we quoted four benchmarks and no results. That is deliberate.

Every hallucination number about a model you could actually pick today lives on a page that rescores itself. Vectara says so in the first paragraph of its own README: "We plan to update this regularly as our model and the LLMs get updated over time." The FACTS Grounding paper says the leaderboard "will be actively maintained over time". That is a virtue in a leaderboard and a trap in an article, because a figure copied on a Tuesday is a claim about Tuesday that keeps being read in March.

The failure is harder to spot than it sounds. A number does not stop being live because it has passed through a vendor's launch chart or an aggregator's summary. Researching this piece, we traced several tidy-looking figures back through that chain and found a live scoreboard at the end of each. In one case a leaderboard's own prose summary had been rewritten inside a day, along with the score it described. A chart citing a scoreboard is a scoreboard.

So the definitions above come from papers and published methodology, which are stable documents with version histories, and no score appears anywhere in this article. You can check the reasoning yourself: open any hallucination leaderboard, write down the top three models and their numbers, and open it again next month.

The habit this leaves you with is simple. When a claim about model accuracy reaches you, find out whether the number came from a paper, which is fixed, or a leaderboard, which is not, and check the date on whichever it was. A vendor's own figure still counts, and it counts as a different kind of claim than an independent measurement.

Check the output, because you cannot check the intent

All of that lands on three habits you can run in your next conversation, whichever chat app you use.

Ask for the source, then open it. Skip "is that right", which invites the model to generate agreement, and ask "which page did that come from", followed by a click. This is the check that catches fabricated citations, and it works because a URL is verifiable in a way confidence is not. A live source helps for anything with a checkable answer. Thirteen of the models in MultiChats can search the live web, and on the web the app decides per message whether to go and look, because Intent Detection is on until you switch it off in Settings, under Tool Preferences. Free accounts get 2 searches per 30 days on a fixed window.

Ask which part of the answer is weakest. "Which claim here would you want checked before someone relied on it" tends to surface the shaky sentence, because you have asked for a list of caveats rather than an answer it must defend. Treat what comes back as a shortlist for your own checking.

Ask for the citation and the answer as two separate turns. A model asked "what is the rule" and a model asked "quote me the passage that states the rule" behave differently, and the gap between the two responses is informative on its own. The second framing forces something checkable into the output, and a quoted passage that turns out not to exist tells you more than any confidence estimate would.

One more move is worth the effort on questions that matter: ask a second model from a different maker and see whether the answers agree. MultiChats runs 59 models from 17 providers under one subscription, so switching mid-conversation or branching a thread takes seconds. Agreement is not proof, since two models can be wrong the same way, and disagreement points straight at the sentence that needs a human.

The short version

No, AI models do not lie, and the reason that matters is practical rather than philosophical. There is no intent to appeal to, so the checks that work on people do not work here, and the ones that do work are all about the output.

A hallucination rate measures something real. It also measures one particular thing, chosen by whoever built the benchmark, and their choice about what a refusal is worth can move a model several places up or down a table without a single weight changing. When you see one of those numbers, the questions that earn their keep are what got counted and on what day. Everything the chart says about which model won rests on those two answers.