Stop calling it hallucination

TL;DR

• “Hallucination” is a psychiatric term for a perception with no external cause. The 2025 OpenAI paper on the subject describes something else: ordinary statistical error, kept alive by benchmarks that reward guessing.
• On SimpleQA, one OpenAI model scored 24 percent accuracy with a 75 percent error rate; a newer one scored 22 percent with a 26 percent error rate. Accuracy-only leaderboards rank the first higher.
• Researchers who need precision already avoid the word. The Nature paper on semantic entropy chose “confabulation” and defined it narrowly: wrong and arbitrary.
• Buyers should ask for the split of correct, wrong and declined answers, the test conditions, and what the product does with a claim it cannot trace to a source.

This is an opinion piece, and the opinion is simple: the industry’s favourite word for a model being wrong is doing harm. “Hallucination” sounds like a diagnosis. It implies a system that normally perceives the truth and occasionally suffers a glitch, and a glitch is something you patch. The published evidence describes a different mechanism, which calls for a different fix. The word used to name the problem shapes which remedies get built, which numbers appear on model cards, and what a buyer thinks to ask.

A word borrowed from psychiatry

The most cited survey on the subject, Ji et al.’s “Survey of Hallucination in Natural Language Generation” in ACM Computing Surveys, is candid about where the term came from. Outside natural language processing, the authors write, “hallucination is a psychological term referring to a particular type of perception”, and they quote a clinical definition: “a percept, experienced by a waking individual, in the absence of an appropriate stimulus from the extracorporeal world”.

The survey then gives the field its working definition, “the generated content that is nonsensical or unfaithful to the provided source content”, and splits it in two. An intrinsic hallucination is “the generated output that contradicts the source content”. An extrinsic hallucination is “the generated output that cannot be verified from the source content”.

Notice what has happened. A clinical word about perception has been attached to two engineering conditions, contradiction and missing support, neither of which involves perceiving anything. The authors chose it because hallucinated text “gives the impression of being fluent and natural despite being unfaithful and nonsensical”. That describes the symptom, not the cause, and the cause is where the word misleads.

What the statistics actually say

In September 2025, OpenAI published a paper titled “Why Language Models Hallucinate”. Its abstract does not describe a glitch. It says that language models “hallucinate because the training and evaluation procedures reward guessing over acknowledging uncertainty”, and that the errors “originate simply as errors in binary classification”.

The argument has two halves. The first is about pretraining. A model learns from text that carries no labels marking statements true or false. Kalai, Nachum, Vempala and Zhang show that generating valid text is at least as hard as classifying whether a statement is valid, so a model inherits as a generator at least the error rate it would have as a classifier of facts. For facts with no pattern behind them, such as birthdays, the error rate is tied to how often the fact appears in training: “if 20% of birthday facts appear exactly once in the pretraining data, then one expects base models to hallucinate on at least 20% of birthday facts”.

The second half is about why post-training does not remove these errors, and it is the half that matters for buyers.

Grading that rewards the guess

Most benchmarks score an answer as right or wrong, with no third box. The paper states the consequence bluntly: “Under binary grading, abstaining is strictly sub-optimal. IDK-type responses are maximally penalized while an overconfident ‘best guess’ is optimal.” Its Table 2 lists GPQA, MMLU-Pro, IFEval, BBH, SWE-bench and Humanity’s Last Exam, and records for each that grading is binary and credit for declining is none. The authors argue this “can only be addressed through a socio-technical mitigation: modifying the scoring of existing benchmarks that are misaligned but dominate leaderboards, rather than introducing additional hallucination evaluations”.

OpenAI’s accompanying post supplies the numbers, drawn from the GPT-5 system card’s SimpleQA results:

On accuracy, the older o4-mini wins by two points. It gets there by answering everything, and three quarters of its answers are wrong. The newer model declines more than half the time and is wrong a quarter of the time. As the post puts it, “accuracy-only scoreboards dominate leaderboards and model cards, motivating developers to build models that guess rather than hold back.”

The proposed remedy is not a new component. It is a grading rule: “Penalize confident errors more than you penalize uncertainty, and give partial credit for appropriate expressions of uncertainty.” The paper suggests writing the confidence threshold into the question itself. That is a change to the exam, not to the student.

A leaderboard that reports only accuracy prefers the model that is wrong three times as often.

The court case that made the word famous

The cost of the wrong word is on record in the sanctions opinion in Mata v. Avianca, issued by the Southern District of New York on 22 June 2023. Two lawyers and their firm, Levidow, Levidow & Oberman, “submitted non-existent judicial opinions with fake quotes and citations created by the artificial intelligence tool ChatGPT, then continued to stand by the fake opinions after judicial orders called their existence into question”. Six cited decisions, the court records, “were generated by ChatGPT and do not exist”. The penalty was $5,000, imposed jointly.

The telling part is the lawyer’s account of what he believed. He had been “operating under the false perception that this website could not possibly be fabricating cases on its own”, and “I just never thought it could be made up.” In a declaration he described the fake cases as having “various indicia of reliability such as case captions, the names of the judges from the correct locations, and detailed fact patterns and legal analysis that sounded authentic.”

That is what the statistical account predicts and the glitch account does not. The model was not misperceiving a database. It was producing the most plausible continuation of a request for case law: a caption, a reporter citation, a judge in the right circuit. Each of those can be generated from patterns. Whether the case exists cannot. The court’s framing was the useful one: “there is nothing inherently improper about using a reliable artificial intelligence tool for assistance”, but the rules “impose a gatekeeping role on attorneys to ensure the accuracy of their filings”. Gatekeeping is a process word. Hallucination is a symptom word. Only one tells you what to do.

What researchers say when they need precision

When Farquhar, Kossen, Kuhn and Gal published their detection method in Nature in June 2024, they chose their vocabulary deliberately. “We focus on a subset of hallucinations which we call ‘confabulations’”, they write, cases where models “fluently make claims that are both wrong and arbitrary”, where arbitrary means the answer changes with irrelevant details such as the random seed. Asked what the drug Sotorasib targets, a model sometimes answers KRAS G12C, which is correct, and sometimes KRAS G12D, which is not, under identical instructions.

Their method, semantic entropy, follows from the definition: sample several answers, group them by meaning rather than wording, and measure how spread out the meanings are. The authors are equally explicit about what the word excludes: being “consistently wrong as a result of being trained on erroneous data such as common misconceptions”, lying “in pursuit of a reward”, and systematic reasoning failures. “Our method explicitly does not directly address” those cases, they write, because they “represent different underlying mechanisms” despite similar symptoms.

The choice was not decoration. A 2017 clinical trial in PLoS ONE by Triviño and colleagues opens with the neuropsychological definition: “A confabulation is the generation of false memories without intention to deceive. Confabulators have a great feeling of rightness and show great resistance to being confronted with the truth.” The condition is studied in brain injury, Korsakoff syndrome and Alzheimer’s disease. A fluent system with no internal flag for “I made this up” is closer to that than to seeing things. It is still a metaphor, which is why Farquhar’s group applied it to one subset only. Precise words build precise detectors.

How vendors measure it today

The better system cards have already moved past the single number. The SimpleQA benchmark, published by OpenAI researchers in November 2024, grades every answer “as either correct, incorrect, or not attempted” and is meant to test whether models “know what they know”.

The GPT-5 system card of August 2025 reports SimpleQA two ways: accuracy, and a “hallucination rate” defined as the share of attempted answers that are wrong. Its Table 8, run without web access, gives gpt-5-thinking 0.55 accuracy and 0.40 hallucination rate against OpenAI o3’s 0.54 and 0.46. The card also reports a production measurement on browsing-enabled ChatGPT traffic, where a model grader with web access checked each claim: 4.5 percent of gpt-5-thinking’s claims were judged wrong against 12.7 percent for o3, and 9.6 percent for gpt-5-main against 12.9 percent for GPT-4o. These are the vendor’s own figures, graded by its own model, which the card says agreed with human raters 75 percent of the time.

Anthropic’s Claude Opus 4.8 system card of May 2026 grades four closed-book sets, including Google’s SimpleQA Verified, as “correct, incorrect, or as an acknowledgement it could not answer the question”, then adds the metric the Kalai paper asked for: “Because a model can inflate its correct-rate by simply guessing on every question, we also report the net score (correct minus incorrect), which penalizes confident wrong answers and rewards well-placed abstention.” Opus 4.8 had the lowest incorrect rate of the six models compared, and got there “mainly by abstaining on questions about which it was uncertain rather than by answering more questions correctly”.

Both vendors now publish the three-way split, and both still call the wrong-answer rate a hallucination rate. The measurement has caught up with the mechanism; the vocabulary has not. Our guide to reading a model card like an auditor covers what else to look for in these tables.

Better words

No replacement is perfect, and this editorial does not propose a single one. It proposes matching the word to the mechanism, because each points at a different fix.

“Confident error” is the phrase this publication will use by default, because the scoring evidence supports it and because it keeps responsibility where the Kalai paper places it: in how the system was graded and what it was allowed to say. “Confabulation” is right when the failure is arbitrary in Farquhar’s sense. “Unsupported claim” is right in any retrieval setting, where the question is whether the source agrees, not the world; our comparison of retrieval and long context is organised around that failure mode.

Questions to ask a vendor

When a vendor says a model hallucinates less, ask these.

1. Show the three-way split: correct, wrong and declined, on questions close to our task. Accuracy alone hides the trade.
2. Under what conditions? The GPT-5 card’s SimpleQA numbers are no-web; its production numbers are browsing-enabled. The two are not comparable.
3. Who graded it? A model, with what agreement against humans, or people, and on how many items.
4. How is a refusal scored in training and evaluation? If declining earns zero and guessing can earn one, the model has been taught to guess.
5. What does the product do with a claim it cannot trace to a source? Flag it, withhold it, or pass it through. The lawyers in Mata v. Avianca had no flag and read its absence as confirmation.

The frameworks that can answer these are the subject of our guide to comparing evaluation tools. The point here is narrower. A confident error is a measurable, scoreable, penalisable thing. A hallucination is an excuse. The models are not seeing things. They are answering when they should decline, because every scoreboard told them to, and changing the scoreboard starts with not calling the result a vision.