AI Hallucination Fact Check: What Viral Videos Get Wrong About LLM Uncertainty

Last Updated: October 2026

Every few weeks a clip goes viral in which someone asks a chatbot a simple question, gets a confident and completely invented answer, and concludes that the model is “lying,” “broken,” or “just autocomplete.” This AI hallucination fact check takes those clips seriously. The clips are usually real, which is exactly why the conclusions drawn from them deserve scrutiny. A fabricated citation is a genuine failure. It does not, by itself, tell you why the failure happened, whether the model “knew,” or whether a newer model, a retrieval layer, or a different prompt would have prevented it.

The science moved a lot in the last twelve months. OpenAI and Georgia Tech researchers argued that hallucination is partly an incentive problem baked into how we grade models. Anthropic published interpretability work showing that refusal is a default circuit that can be switched off by mistake. Independent leaderboards now measure hallucination separately from accuracy, and September 2026 brought a frontier launch, GPT-6 Astra, whose reported gains centered on lower hallucination rather than only a higher score.

What this covers: a fair scoring of six viral claims against primary sources, the mechanism behind confident errors, why “does the model know it is wrong” has a more interesting answer than yes or no, what the current benchmarks actually measure, and a layered engineering checklist for reducing false claims in production.

What Changed for October 2026

This is a ground-up rewrite of the April 2026 version of this article. If you read the earlier post, treat its specifics as superseded.

The explanation matured. The 2025 paper “Why Language Models Hallucinate” (Kalai, Nachum, Vempala and Zhang) gives a formal account: pretraining errors follow from the statistics of binary classification, and benchmarks that score only accuracy reward guessing over honest abstention. The April version leaned on informal explanations such as “spurious correlations.” Those are not wrong, but they are no longer the best account.
Interpretability gave a mechanism. Anthropic’s circuit-tracing work describes a default “decline to answer” pathway that a “known entity” feature can inhibit, and shows how a misfire of that feature produces a fabricated answer.
Hallucination is now measured on its own axis. Artificial Analysis’s AA-Omniscience benchmark reports accuracy, hallucination rate, and a combined index that penalizes wrong answers while leaving refusals neutral. Vectara’s leaderboard tracks grounded summarization separately.
A frontier launch put hallucination front and center. Artificial Analysis reported on September 9, 2026 that GPT-6 Astra cut its AA-Omniscience hallucination rate from 92% to 51% at maximum effort, while accuracy rose by 4 points. A hallucination rate of 51% is still high, and the next sections explain why.
Some earlier figures were removed. The April post quoted several precise percentages without a traceable source. They are gone. Every number below links to a primary source or is labelled as secondary reporting.
The viral claims themselves changed. “AI always makes things up” is now often replaced by its opposite, “hallucination is solved.” The evidence supports neither.

Context and Background

“Hallucination” is a loose word, and much of the confusion in viral videos is definitional. In the research literature it covers at least three different failures. The first is a factual error with no source in view, such as inventing a court case or misattributing a quote. The second is unfaithfulness to provided context, such as a summary that adds a claim the document never made. The third is fabricated self-reports, such as claiming to have run code or checked a source when no tool call happened.

These have different causes and different fixes. A retrieval system helps with the first and gives the model a new way to fail at the second. Neither addresses the third, which is closer to an honesty and training-signal problem. When a video blends them, a fix that works for one looks like a fix that “didn’t work.”

A second source of confusion is that the word suggests perception. A human hallucination is a sensory experience without a stimulus. A language model has no stimulus-and-percept pipeline to malfunction. Some researchers prefer “confabulation,” a term from neuropsychology for fluent, arbitrary false statements produced without intent to deceive. The Nature paper on semantic entropy adopts exactly that framing, and it is a better fit for the cases that matter in practice.

The most-cited legal example remains Mata v. Avianca (S.D.N.Y., 2023), in which attorneys filed a brief containing case citations that a chatbot had invented and the court imposed sanctions. It is a useful anchor because it shows the real-world cost pattern: the error was fluent, formatted like a real citation, and unverified by the humans downstream. For more on how evaluation fits into a production pipeline, see our guide to LLM evaluation with LLM-as-judge pipelines.

The rest of this article uses a simple standard. A claim is supported if primary research or a reproducible benchmark backs it, partly true if it describes a real phenomenon but draws the wrong conclusion, and unsupported if the evidence points the other way or does not exist. Our sibling fact-check on the larger “is AGI here” question applies the same method: see AGI in 2026: viral videos versus what frontier models do.

Why Language Models Hallucinate: The Mechanism Behind the Clips

A language model is trained to predict the next token given the previous ones. That objective rewards producing text that is statistically plausible, and it has no built-in notion of “I should stay silent here.” The direct answer to why models hallucinate is that fluent completion is what they are optimized for, and two separate pressures then keep false completions alive: pretraining cannot always tell valid from invalid statements, and post-training and benchmarks reward guessing over abstaining.

Figure 1: Two pressures behind hallucination. Pretraining leaves a gap between plausible and true statements, and binary scoring then rewards filling that gap with a guess.

The figure traces the argument from the OpenAI paper. The left branch is about knowledge: some facts are learnable from patterns, and others are not. The right branch is about incentives: once a model is deployed against benchmarks that score only correct answers, guessing dominates abstaining. The sections below take each in turn.

Pressure one: errors are a statistical consequence of pretraining

Kalai, Nachum, Vempala and Zhang frame hallucination as a binary classification problem in disguise. Imagine asking, for every possible sentence, “is this valid?” A model that generates text well must implicitly be able to separate valid outputs from invalid ones. When the training data does not contain enough signal to make that separation, errors are not a bug in the optimizer. They are a lower bound set by the statistics.

The paper’s clearest illustration is the arbitrary fact. Consider a person’s birthday. If a given individual appears in the training data exactly once, there is no pattern from which to infer the date. The model sees the fact, cannot generalize it, and has no way to tell a remembered birthday from a plausible one. The authors relate the expected error rate on such facts to the fraction of facts that appear only once in the corpus, an idea borrowed from the Good-Turing estimator of missing mass. You do not need the proof to see the intuition: rare facts are the ones models cannot reliably separate from plausible fiction.

This explains a pattern the viral clips illustrate by accident. Chatbots are strong on famous entities and weak on obscure ones. Ask about a global figure and the model has seen thousands of consistent mentions. Ask about a mid-level academic, a small town’s mayor, or a niche product, and you are in the long tail where single-occurrence facts live.

The paper also notes that other classes of error follow from the model family rather than the data. A model with poor representation of a task, or a computationally hard problem, will fail regardless of how much data it saw. The authors list both as additional sources, alongside distribution shift and errors in the training data itself (“garbage in, garbage out”). A related theoretical result by Xu, Jain and Kankanhalli argues that a computable language model cannot avoid hallucination on all inputs under a formal definition, which supports the same practical takeaway: expect a floor, not a zero.

Pressure two: post-training and benchmarks reward guessing

If pretraining explains why errors exist, the incentive argument explains why they survive. Most widely used benchmarks give one point for a correct answer and zero for a wrong answer or a blank one. Under that rule, a model that always guesses scores at least as well as one that says “I don’t know,” and usually better. The OpenAI write-up compares it to a student taking a multiple-choice exam with no penalty for wrong answers: the rational strategy is to fill in every bubble.

The paper backs this with an example from OpenAI’s own SimpleQA evaluation. Reported figures for two models:

Metric (SimpleQA)	gpt-5-thinking-mini	OpenAI o4-mini
Abstention rate	52%	1%
Accuracy rate	22%	24%
Error rate	26%	75%

The older o4-mini edges ahead on raw accuracy, 24% to 22%, yet it is wrong three times as often. A leaderboard that ranks only on accuracy would place the more error-prone model first. That is the incentive problem in one table. These numbers come from OpenAI’s own post and are on a single, deliberately difficult benchmark, so they show direction, not a general hallucination rate for either model.

The authors’ remedy is not another hallucination benchmark. They argue the dominant evaluations should change their scoring so that confident errors cost more than abstentions, and appropriate expressions of uncertainty earn partial credit. In their words, the primary eval metrics need to be reworked. That is a social change as much as a technical one, because leaderboards shape what labs optimize.

Pressure three: a circuit, not a database

The third piece is mechanistic. Anthropic’s interpretability team traced how Claude decides whether to answer a factual question and found that refusal is the default. In their description, a circuit is “on” by default and causes the model to say it has insufficient information. When the model recognizes an entity, a competing “known entities” feature activates and inhibits the default circuit, so the model proceeds to answer.

Hallucination, in this account, is a misfire. If the known-entity feature activates for a name the model has only faint familiarity with, it suppresses the decline circuit even though the underlying knowledge is not there. The model then does what it was trained to do: it produces a plausible continuation. Anthropic demonstrated this by getting Claude to assert that an invented person, “Michael Batkin,” plays chess.

Figure 2: Default decline versus known-entity inhibition. The same inhibition that enables correct answers also enables confabulation when the familiarity signal misfires.

This reframes the viral claim “the model doesn’t know that it doesn’t know.” The picture is more nuanced. There is a recognition signal, and it is separate from the retrieval of the actual fact. When the two come apart, you get a confident wrong answer. Anthropic’s caveat applies: this is one analysis of a specific model family on specific prompts, and circuits described this way are simplified views of a large network.

Why this matters for how you read a clip

Put the three pressures together and the typical viral demo becomes predictable. The prompt targets the long tail (an obscure person, a niche paper, a fictional place), the benchmark-trained instinct is to answer, and the familiarity signal misfires on a plausible-sounding name. None of that requires deception or malfunction. It is the system working as trained on a question it should have declined.

That is also why the fix cannot be a single trick. Pretraining statistics set a floor. Scoring rules set an incentive. Inference-time checks, retrieval, and verification can reduce the rate at which the first two reach the user, and the later sections cover them. For a current frontier model profile, see our Claude Opus 5.5 explainer.

Six Viral Claims, Scored

Here is the claim-by-claim scorecard. Each verdict applies the standard from earlier: supported, partly true, or unsupported.

#	Viral claim	Verdict	One-line reason
1	“The model is just making things up”	Partly true	Confabulation is real, but it is concentrated in the long tail and varies by model and setting
2	“The model knows it is lying”	Unsupported	No evidence of intent, though internal signals about familiarity do exist
3	“Newer models hallucinate more, so AI is getting worse”	Partly true	Some releases did regress on some measures, but the trend across recent flagships is mixed, not monotonic
4	“Hallucination is solved”	Unsupported	Even the best current benchmark numbers are far from zero
5	“Temperature zero removes hallucination”	Unsupported	It removes sampling variance, not wrong high-probability answers
6	“RAG fixes it”	Partly true	Grounding helps a lot, and creates new failure modes

Claim 1: “The model is just making things up”

The observation is accurate and the framing is lazy. Models do generate false statements fluently, and the OpenAI paper explains the mechanism. What the clip omits is the distribution of errors. Hallucination rates depend heavily on what you ask. Questions about famous, frequently documented facts are answered well. Questions about rare entities, precise numbers, recent events, and citations are where models fail, and benchmarks such as SimpleQA were built specifically by selecting questions that GPT-4 answered incorrectly. Generalizing from an adversarially chosen prompt to “the model always makes things up” is a sampling error.

There is also a measurement point. Anyone who has seen a viral clip has seen a selected failure, and the denominator is invisible. A serious evaluation reports a rate over a fixed question set. The question to ask of any clip is: how often does this happen on a representative sample of this question type, and for which model version and settings?

Claim 2: “The model knows it is lying”

Lying requires an intent to create a false belief. There is no evidence that a standard language model forms such an intent when it confabulates. The Anthropic analysis points the other way: the model’s decline circuit was overridden by a misfiring familiarity signal, and the answer that followed was plausible text, not concealed truth.

A more careful claim is that models carry internal information about their own reliability. Kadavath and colleagues at Anthropic reported in 2022 that large models can be reasonably well calibrated when asked to evaluate whether their own answers are true, and that a model can be trained to predict “P(IK)”, the probability that it knows the answer. This does not mean the model is lying when it errs. It means the uncertainty signal exists but is not reliably routed into the output. That is the gap that abstention-aware training tries to close.

One caution cuts the other way. Models can and do produce false statements about their own actions, such as claiming to have run a test or opened a link. That is a distinct failure from factual confabulation, and OpenAI’s launch material for GPT-6 Astra reports measuring it separately as a “capability hallucination” rate. The reported drop, from 9.4% for GPT-5.6 Sol to about 2% for Astra at long solution lengths, comes from a secondary summary of OpenAI’s materials and was measured on de-identified conversations that users had flagged as containing factual errors. Those conversations were deliberately hard, so the figures show relative improvement and are not a production error rate.

Claim 3: “Newer models hallucinate more”

This claim has a real kernel. In April 2025, OpenAI’s system card for o3 and o4-mini reported PersonQA hallucination rates of 16% for o1, 33% for o3 and 48% for o4-mini, with GPT-4.5 at 19% and GPT-4o at 30% (figures as relayed by Techzine). OpenAI said the underlying cause was unknown and that o3 “generally makes more assertions,” which yields both more correct and more incorrect claims. That is a regression on one measure for one model family, and it also shows the abstention trade-off at work.

What can be verified is the opposite direction. The SimpleQA table above shows a newer model abstaining far more and erring far less than an older one. Artificial Analysis reported that GPT-6 Astra reduced its AA-Omniscience hallucination rate from 92% to 51% at maximum effort while raising accuracy. The honest summary is that hallucination is not monotonic across releases. It depends on the model, the reasoning budget, the benchmark, and whether the lab explicitly trained for abstention. “Newer means worse” and “newer means solved” are both overclaims.

Claim 4: “Hallucination is solved”

The mirror-image claim is now common in launch-week commentary. The evidence does not support it. On AA-Omniscience, whose hallucination rate is defined as the share of non-correct responses that were incorrect rather than abstentions, the headline improvement for GPT-6 Astra still leaves the metric at 51% at maximum effort. That means that on questions the model did not get right, roughly half the time it answered anyway. Meanwhile the lowest hallucination rate on the same benchmark’s page belongs to a 1-billion-parameter model, MiniCPM5-1B, at 1%. That is not a sign of strong factual knowledge. A tiny model that refuses almost everything has a very low hallucination rate and very low accuracy. It is why you need both axes.

Theory points the same way. The OpenAI authors state that accuracy will never reach 100%, because some real-world questions are inherently unanswerable. A system that never errs would have to know when to stay silent, which is an uncertainty problem, not a knowledge problem.

Claim 5: “Temperature zero removes hallucination”

Temperature rescales the probability distribution before sampling. At zero, the decoder picks the highest-probability token every time, which makes output more repeatable. It does not alter which token is highest. If the most probable continuation is a fabricated name, greedy decoding returns that fabrication every time, with perfect consistency. Consistency and truth are different properties.

Two details are worth adding. First, temperature zero is not guaranteed to be bit-for-bit deterministic in practice, because batch composition and floating-point reduction order can change results on shared inference servers. Second, there is a real inverse effect that the semantic entropy method exploits: when a model is unsure, repeated samples at nonzero temperature disagree, and that disagreement is a signal. Greedy decoding hides it.

Claim 6: “RAG fixes it”

Retrieval-augmented generation puts relevant documents in the prompt so the model can answer from text instead of from parametric memory. For questions that a document answers, this is the single most effective mitigation in common use, and the Vectara leaderboard exists precisely because even grounded summarization is not perfectly faithful. The failure modes are well-known in practice: the retriever misses the right passage and the model answers from memory anyway, the model blends two retrieved passages, or it paraphrases a number incorrectly. These are covered in depth in our posts on corrective and self-reflective RAG and agentic RAG retrieval agents. The summary: RAG moves the problem from “does the model know?” to “did we retrieve it, and did the model stay faithful to it?”

Uncertainty Is Measurable, but Not by Asking “Are You Sure?”

The most common home remedy shown in viral clips is to ask the model whether it is certain. Sometimes the model confidently doubles down, and sometimes it caves immediately and apologizes for a correct answer. Both behaviors are demonstrated and neither is a measurement. Verbalized confidence is just more generated text, and it inherits the same training pressures as everything else the model says.

There are better signals, and they sit on a spectrum from cheap to expensive.

Token probabilities and their limits

The most direct signal is the probability the model assigned to its own output. In theory, a calibrated model assigns 90% to answers that are right 90% of the time. Kadavath and colleagues found that large pretrained models are reasonably calibrated on multiple-choice formats. Post-training changes that picture: OpenAI’s GPT-4 technical report observed that the base model’s calibration was good and that post-training degraded it. I am citing that from the report’s published discussion and have not re-verified the figure in this session, so treat it as directional.

The deeper problem is that token probability confuses two things. A model may be unsure which words to use (many phrasings of the same answer) or unsure which fact is true. Raw sequence likelihood penalizes the first kind of uncertainty, which is harmless, and so is a noisy hallucination detector.

Semantic entropy: measuring uncertainty over meaning

Farquhar and colleagues addressed that confusion in a 2024 Nature paper. Their method samples several answers to the same prompt, groups them by meaning using bidirectional entailment (two answers belong together if each implies the other), and computes entropy over the meaning clusters rather than over word sequences. If ten samples collapse into one cluster, the model is stable. If they scatter across many clusters, the model is guessing differently each time.

Figure 3: Semantic entropy for LLM uncertainty. Sample many answers, cluster by meaning, and treat high entropy as a flag to abstain or verify.

The authors report an AUROC of 0.790 for detecting incorrect answers, against 0.691 for naive entropy, 0.687 for embedding regression and 0.698 for the P(True) baseline, across several open model families. AUROC is a rank-based score where 0.5 is chance and 1.0 is perfect, so 0.79 is useful but far from infallible. The authors are explicit about the boundary: the method targets confabulations, meaning arbitrary and inconsistent errors. It does not help when a model is systematically wrong, for example when it confidently repeats a widespread misconception. Consistent errors have low entropy.

The practical costs also matter. Sampling N answers multiplies inference cost by roughly N, plus an entailment pass to cluster them. That is acceptable for a high-stakes answer and prohibitive for every chat turn, which is why production systems use it selectively.

Calibration is a training choice

A model can be trained to say “I don’t know” at the right times, and that is what the incentive argument predicts. The SimpleQA data in the OpenAI post show a reasoning model abstaining on 52% of questions where an older one abstained on 1%. Whether that is the right trade depends on the application. A medical triage tool wants high abstention. A brainstorming tool wants low abstention and lots of speculation. This is a product decision, and a model with a single global abstention setting will disappoint someone.

What the Benchmarks Actually Measure

Viral comparisons often cite a “hallucination rate” without saying which one. Four widely used sources measure different things and cannot be compared directly.

Benchmark	What it measures	Unit of failure	Key caveat
SimpleQA (OpenAI, 2024)	Short fact-seeking questions with one indisputable answer	Incorrect vs not attempted	Questions were adversarially collected against GPT-4 responses, so absolute scores are pessimistic
AA-Omniscience (Artificial Analysis)	About 6,000 questions across 6 domains and 42 topics	Share of non-correct answers that were wrong rather than abstained	Rate alone rewards refusing; read it with accuracy
Vectara HHEM leaderboard	Factual consistency of summaries against source documents	Summary claims unsupported by the document	Measures grounded summarization, not general knowledge
PersonQA (OpenAI internal)	Questions about people from a fixed set of facts	Incorrect answers	Internal eval, narrow domain

Three observations follow.

The same model can look excellent on one and poor on another. The Vectara GitHub page, last updated May 11, 2026 at the time I read it, listed Antgroup Finix S1 32B at 1.8%, OpenAI GPT-5.4-nano at 3.1% and Google Gemini 2.5 Flash Lite at 3.3% as leaders. That dataset uses more than 7,700 articles, is not public, and is scored by the HHEM-2.3 model at temperature zero where possible. Those low single-digit rates describe summarizing a provided document, which is a far easier task than answering open-book trivia without one. Third-party blog tables of the same leaderboard show different figures for other models, which is typical when snapshots are taken on different dates or a harder dataset version is used.

A combined index hides the trade-off, so look at both axes. AA-Omniscience’s index runs from negative 100 to 100, rewards correct answers, penalizes hallucinations and does not penalize refusals. On the page I read, Claude Opus 5.5 led the index at 46, with GPT-6 Astra at its High setting at 44, and Claude Fable 5.1 had the highest accuracy at 67%. An index like this moves when a model gets more knowledgeable and also when it gets more careful, and you cannot tell which from the headline.

Benchmarks decay. Public question sets leak into training data, vendors tune against them, and a score on a saturated test stops separating models. That is why Vectara keeps its set private and why serious teams build a held-out evaluation from their own domain. Our post on LLM-as-judge evaluation pipelines covers how to do that without fooling yourself.

What Actually Reduces False Claims in Production

No single control removes hallucination, so reliable systems stack several, each catching what the previous layer lets through. The reference pattern below is the one that most deployments converge on: retrieve, generate with citations, verify each claim against the sources, and abstain or escalate when support is missing.

Figure 4: A layered defense against LLM hallucination. The retrieval miss path, shown on the right, is the silent failure that most deployments underestimate.

Layer 1: constrain what the model must answer from

Ground the answer in retrieved text and instruct the model to answer only from it. Require citations that point to passage identifiers, not free-text references. This changes the task from recall to reading comprehension, which models do better. It also makes verification possible, because a claim now has a claimed source. For the trade-offs between putting knowledge in weights, in retrieval and in long context, see our comparison of fine-tuning, RAG and long-context approaches.

Layer 2: verify claims, not answers

Split the draft into atomic claims and test each against the cited passage with an entailment model or a second LLM acting as judge. Claims that are unsupported get removed or rewritten. Vectara’s HHEM is exactly this kind of classifier, released in an open variant (HHEM-2.1-Open), which makes it a practical building block. A verifier is itself fallible, so measure its precision and recall on a labelled sample of your own traffic and do not assume the vendor figure transfers.

Layer 3: make abstention a first-class output

Give the system a legal way to say “I could not find support for that.” Route low-confidence cases to a human or to a narrower fallback. Semantic entropy or agreement across a few samples can serve as the trigger for high-stakes queries. Pair this with structured outputs and validation so downstream code can distinguish an answer from a refusal; see our guide to LLM output validation and guardrails.

Layer 4: measure continuously

Hallucination rates drift when models, prompts, retrieval indexes or user populations change. Log the retrieved passages, the claims and the verifier verdicts. Track an unsupported-claim rate and an abstention rate together, because optimizing one alone will degrade the other. Our post on LLM observability and LLMOps architecture covers the tracing side.

A worked example with illustrative numbers

The numbers here are illustrative, not measured. Suppose a support assistant answers 1,000 questions. Without controls, 10% contain at least one unsupported claim, so 100 answers are wrong in some way. A claim verifier catches 70% of those, leaving 30. An abstention rule triggered by retrieval confidence withholds some answers: say it blocks 5% of all answers, of which half would have been among the bad ones. The result is that user-visible bad answers fall from 100 to roughly 5 to 25 depending on the overlap, while 25 to 50 good answers are withheld. The point is the shape of the trade, not the figures: every reduction in false claims is bought with some abstention or latency, and you should choose the exchange rate deliberately.

Trade-offs, Gotchas, and What Goes Wrong

Optimizing the wrong metric. Teams that chase a low hallucination rate alone will find that the cheapest way to get it is to refuse more. Always report the pair: accuracy and hallucination rate, or answered-correct and answered-wrong with abstentions as a third bucket. The 1% hallucination rate held by a 1-billion-parameter model on AA-Omniscience should be a permanent reminder.

Silent retrieval failures. When retrieval returns nothing relevant, many pipelines still call the generator, which then answers from its weights with the confident tone of a grounded system. Users cannot see that no source was used. Detect low retrieval scores and treat them as an abstention trigger.

Verifier shared blind spots. If the verifier is the same model family as the generator, they can share misconceptions. Systematically wrong beliefs, the kind semantic entropy cannot detect, can pass both checks. Use a different model family or an external knowledge source for high-stakes checks.

Citations as theater. A formatted citation is not evidence of grounding. Models can attach a real-looking reference to a claim the source does not support. Verify that the cited passage actually entails the claim.

Overcorrection. Systems tuned to avoid any possible error become unusable, and users route around them. The right target depends on cost of error: a legal research tool and a recipe suggester should not share an abstention threshold.

Unfalsifiable self-reports. A model’s explanation of why it answered as it did is a generated rationale, not a readout of its computation. Interpretability work such as Anthropic’s circuit tracing exists precisely because the model’s own account is not trustworthy evidence about its internals.

How to Fact-Check the Next Viral Clip

A short procedure keeps you from being fooled in either direction.

Identify the failure type. Is it a factual error, unfaithfulness to a provided document, or a false claim about the model’s own actions?
Check the model and settings. Which model version, with search or tools on or off, and at what reasoning effort? Results differ widely across these.
Check the question type. Is it a long-tail fact, a citation, or a trick question about something fictional? These are the hardest prompts by design.
Ask for the denominator. One failure is an anecdote. Look for a rate on a fixed question set.
Reproduce it. Run the prompt several times. If answers differ, you are looking at confabulation. If it is stable and wrong, you are looking at a systematic error.
Check whether the model could have abstained. If the right answer was “I don’t know” and it guessed, the incentive account applies.

Practical Recommendations

If you build with language models, treat hallucination as a managed risk with a budget, not as a defect to be eliminated. Start by deciding the cost of a wrong answer versus the cost of a withheld one in your domain, because that decision sets every threshold downstream.

Choose models by the pair of numbers that matters to you. For open-domain assistance, read accuracy and hallucination rate together on a benchmark close to your task, then build a held-out test from your own data. For document-grounded work, look at grounded-summarization results and run your own faithfulness checks.

Prefer systems that can abstain, and test the abstention path explicitly. Give users visible citations, and spot-check that they support the claims. Keep a human in the loop where an error is expensive or irreversible.

Checklist:

Define acceptable error and acceptable abstention rates per use case.
Ground answers in retrieved text and require passage-level citations.
Verify claims against sources with a verifier you have measured yourself.
Add an abstain or escalate path and test it.
Treat retrieval misses as a first-class failure signal.
Track unsupported-claim rate and abstention rate together, over time.
Re-run your held-out evaluation on every model or prompt change.

Frequently Asked Questions

What is an AI hallucination?

An AI hallucination is a fluent statement from a language model that is false or unsupported by the source it should rely on. Researchers distinguish factual errors, unfaithfulness to provided documents, and false claims about the model’s own actions. Some prefer the term confabulation for arbitrary false statements produced without intent. The common thread is that the output sounds plausible, which is what makes it hard to catch without checking against a source.

Why do language models hallucinate?

Two reasons work together. During pretraining, a model cannot reliably separate true from plausible statements for rare facts it saw only once. Then standard benchmarks score a guess the same as or better than an honest “I don’t know,” so training favors guessing. OpenAI’s 2025 paper makes both arguments formally. Anthropic’s circuit analysis adds that a misfiring familiarity signal can switch off the model’s default decline behavior.

Can hallucinations be eliminated completely?

No current evidence says so. The OpenAI authors argue accuracy will never reach 100% because some questions are unanswerable, and formal results suggest a floor for any computable model. In practice, error rates can be reduced a great deal with grounding, verification and abstention. Even a best-in-class launch such as GPT-6 Astra still sits at a 51% AA-Omniscience hallucination rate at maximum effort, per Artificial Analysis.

Does setting temperature to zero stop hallucinations?

No. Temperature zero makes the decoder pick the most likely token each time, which improves repeatability. If the most likely answer is wrong, you get the same wrong answer every time. Zero temperature can also hide a useful warning sign, because disagreement across sampled answers is how methods such as semantic entropy detect uncertainty. It is also not perfectly deterministic on shared inference infrastructure.

Does RAG eliminate hallucination?

No, but it is among the most effective mitigations. Retrieval lets the model answer from supplied text instead of memory. New failures appear: the retriever may miss the right passage, the model may blend passages or misquote numbers, and a citation can be attached to a claim the source does not support. Verify claims against cited passages and treat low retrieval scores as a reason to abstain.

Which AI model hallucinates the least?

It depends on the benchmark, and a single winner would be misleading. Document-grounded summarization, open-domain trivia and long agentic tasks produce different rankings, and a model that refuses more will show a lower hallucination rate with lower accuracy. Read accuracy and hallucination rate together, check the benchmark date, and test candidates on your own data before choosing.