AI hallucinations: what they are, why chatbots make things up, how often it happens, and a 5-step fact-check for any AI answer

Disclosure: NOXI publishes this guide and is one of the tools it mentions; the fact-check method works with any AI assistant and any presentation software.

An AI answer sounds equally sure when it is right and when it is invented, so the tone tells you nothing. What works is a habit you can time: about two minutes per checkable claim, and a source you opened yourself. The rule in one line: an AI answer is a draft, not a source. Find the original for every number, name, date, quote and citation, read it yourself, and delete whatever you cannot find.

Key numbers

What are AI hallucinations?

AI hallucinations are outputs from a generative AI system that are presented confidently but are false, unsupported or made up: invented facts, wrong figures, quotes nobody said, or references to sources that do not exist. The US National Institute of Standards and Technology (NIST) uses the term confabulation for the same thing [8].

OpenAI's help article on whether ChatGPT tells the truth (opened 22 October 2026) gives three typical forms: incorrect definitions, dates or facts; fabricated quotes, studies, citations or references to non-existent sources; and overconfident answers to ambiguous or complex questions [10]. Anthropic's help centre says Claude can "write things that might look correct but are very mistaken" [11], and Google's Gemini Apps Help says Gemini "can hallucinate and present inaccurate information as factual" [12]. All three vendors document the problem in their own help centres.

If you need the one-line answer for a quiz: AI hallucinations refer to confident AI outputs that are false or not supported by any real source.

Confabulation vs hallucination

Confabulation and hallucination name the same failure; confabulation is the term NIST uses in its Generative AI Profile (NIST AI 600-1, July 2024), and hallucination is the everyday word. NIST notes that some commenters object to "hallucination" and "fabrication" because they attribute human characteristics to software [8].

The second half matters in practice: an answer can ignore the document you supplied, or contradict itself, without inventing a single fact. Check consistency as well as facts.

AI hallucination in images

In image generation, hallucination means a picture that shows something that is not there or is wrong, such as unreadable text on a sign, a building with invented features or a chart with numbers that match nothing. Check a generated image the way you check a sentence: against a real photograph, map or dataset.

Why does AI hallucinate?

AI hallucinates because a language model generates the most likely next words, not the verified ones. NIST describes confabulations as "a natural result of the way generative models are designed": they produce outputs that approximate the statistical patterns of their training data, which can be accurate and can also be false [8].

Three causes explain most of what you will see.

1. Prediction, not lookup. A model without a search tool answers from patterns learned in training. NIST notes this is particularly risky for open-ended, long-form answers and for questions that need specialist knowledge [8].
2. Facts the model saw rarely. The OpenAI paper "Why Language Models Hallucinate" (Kalai, Nachum, Vempala and Zhang, 4 September 2025) argues that hallucinations start as ordinary classification errors. For arbitrary facts with no pattern, such as birthdays, it gives a lower bound: if 20% of birthday facts appear exactly once in the training data, a base model can be expected to hallucinate on at least 20% of birthday facts [9].
3. Tests reward guessing. The same paper argues that hallucinations survive later training because most benchmarks grade answers right or wrong, with no credit for "I don't know". A model that guesses scores higher than one that admits uncertainty, so models are, in the authors' words, always in "test-taking" mode [9].

A fourth cause is time. OpenAI's help article lists the knowledge cutoff: a model trained on data up to a point does not know later events unless it uses a tool such as search [10]. Anthropic's help article gives the same example: Claude "may get confused when prompted about current events" [11].

The second row uses the scoring rule the OpenAI paper proposes: answer only if you are more than t confident, because a mistake costs t ÷ (1 − t) points [9]. At t = 0.9 a mistake costs 0.9 ÷ 0.1 = 9 points. You can borrow the same rule in a prompt (Prompt 2 below).

Why does AI make up sources and references?

AI makes up sources because a citation is just more text to predict: author names, a plausible title, a year and a journal that fit the pattern of a real reference. Unless the assistant actually retrieved the document, nothing in the generation step checks that the reference exists.

Sourcing is also where answers failed most often in the EBU/BBC study: 31% of news answers had significant sourcing problems, including information not supported by the cited source and incorrect or unverifiable sourcing claims [1]. A link under an answer tells you where to start checking; it does not tell you the claim is on that page.

How often does AI hallucinate?

How often AI hallucinates depends on the task, the test and the date, so there is no single hallucination rate. Each study below measures something different, on a different sample and date, so its figure stands alone: the rows cannot be averaged or read as one scale.

The spread is the lesson: summarising a document you supplied is a constrained task, answering knowledge questions from memory is not, and news questions depend on fresh sources the assistant must find and represent correctly. Different tasks, different failure rates.

Two arithmetic checks keep the table honest. The BBC subset improved by 51 − 37 = 14 percentage points, which is 14 ÷ 51 = 27.5% fewer answers with a significant issue in relative terms. Documented AI incidents rose by 362 − 233 = 129, and 129 ÷ 233 = 55.4%; that count covers every kind of AI harm, so it is not a hallucination rate.

AI hallucination rate by model (2026)

A hallucination rate by model is only meaningful for one named benchmark on one date, because models change monthly and each benchmark tests a different task. Read any leaderboard with three questions: who ran it, what task it measured, and when it was last updated.

The 2026 AI Index reports its 22% to 94% range on one accuracy benchmark [3]. Vectara's leaderboard measures summarising and is run by a company that sells hallucination detection [14]. FACTS is published by Google DeepMind, whose own models are among those scored [4]. None of them is a ranking of which assistant you should trust, and this guide does not rank vendors on any one of them.

AI hallucinations examples

The most common AI hallucinations are fabricated citations, outdated facts, wrong numbers, misattributed quotes, sources that do not say what the answer claims, and false statements an assistant makes about itself. The documented examples below come from court judgments, the EBU/BBC report and vendors' own help pages.

The OpenAI paper opens with a small example of its own: asked for one of the authors' birthdays, and told to answer only if it knew, an open-source model gave three different wrong dates on three attempts [9]. Ask the same factual question twice; if the answers differ, at least one is invented.

AI hallucinations in legal cases

AI hallucinations in legal cases usually take the form of fake case law: citations to judgments that do not exist, or real judgments quoted for passages they do not contain. Courts treat the filing, not the tool, as the problem: the lawyer who signs is responsible for every citation.

England and Wales, 6 June 2025. In Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin), the Divisional Court dealt with two referrals together [6]. In the first, a wasted costs application said five fake cases had been cited. In the second, a court schedule listed 45 citations; in 18 the case cited did not exist, so 18 ÷ 45 = 40.0% were invented outright, and many of the real ones did not contain the quoted words. The court said freely available generative AI tools "are not capable of conducting reliable legal research", and that lawyers who use them have a professional duty to check the research against authoritative sources: the government's legislation database, the National Archives' judgments database, the official Law Reports and reputable legal publishers [6]. It decided not to start contempt proceedings in the first case, said that decision "is not a precedent", and referred the lawyer to the regulator.

United States, 3 June 2026. In Lnu v. Blanche, No. 24-4790, the Ninth Circuit sanctioned two attorneys whose briefs contained nonexistent cases, misattributed quotations and misrepresentations of real cases [7]. The court staff's summary published with the order says the rules are broken "at the point of signing and filing", not at the point of using AI, and that the six-month suspension from practice before the court, with monetary sanctions, was owed to a repeated failure of candour about where the errors came from. The order's advice to the bar is a fact-check rule in one line: read everything cited in a court filing, whether drafted by generative AI or not [7].

Worldwide, 19 October 2026. Damien Charlotin's AI Hallucination Cases database listed 2,202 decisions [5]. It mainly counts cases where a court or tribunal found, explicitly or implicitly, that a party relied on hallucinated content, with a few where AI use was alleged but not confirmed; it is a researcher's tracker, not an official register.

Can AI be wrong when it sounds sure?

Yes: AI can be wrong while sounding completely sure, because its tone is generated the same way as its content. OpenAI's help article puts it plainly: "Confidence isn't reliability", and the model may express high confidence in incorrect answers [10]. NIST adds that the risk arises when people believe false content "often due to the confident nature of the response" [8].

Two findings show why fluency misleads. The OpenAI paper notes that "bluffs are often overconfident and specific", such as "September 30" rather than "sometime in autumn" [9]. And the 2026 AI Index reports that models handle a false statement well when it is presented as another person's belief, but that performance "collapses" when the same false statement is presented as something the user believes [3]. In plain terms: if your question contains a wrong assumption, the answer may agree with you.

Specificity is not evidence: a precise figure or a citation with volume and page numbers is exactly what a confident guess looks like. Ask neutral questions rather than leading ones, and treat agreement with your own premise as a reason to check.

How accurate is ChatGPT (and Gemini, Claude)?

No single accuracy figure exists for ChatGPT, Gemini or Claude; accuracy depends on the version, the tools switched on and the task. What exists is two kinds of evidence: what each vendor documents about its own assistant, and independent studies that measured one task at one time. For plans and tools side by side, see ChatGPT vs Claude.

What the vendors say

• OpenAI (ChatGPT): ChatGPT "can produce incorrect or misleading outputs"; tools such as search and deep research give cited answers; the advice is to use it "as a first draft, not a final source" and to verify quotes, data and references [10].
• Anthropic (Claude): users "should not rely on Claude as a singular source of truth"; with web search, review Claude's cited sources, because the original pages may hold context the summary left out (help article dated 17 March 2026) [11].
• Google (Gemini): Gemini Apps "may sometimes produce inaccurate" information; do not rely on responses as medical, legal, financial or other professional advice; double-check responses about people [12]. When sources are available, a Sources button appears at the bottom of the response or in line [13].

ChatGPT accuracy rate and Gemini accuracy in one independent study

The EBU/BBC study reported figures per assistant for one task: free versions answering news questions, late May to early June 2025 [1]. The share of answers with at least one significant issue was 36% for ChatGPT and 76% for Gemini; on accuracy alone, which covers facts and direct quotes, all four assistants tested fell between 18% and 22%. The report attributes most of Gemini's figure to sourcing: 72% of its answers had a significant sourcing issue.

Read these numbers as what they are: one study of free versions, on news, more than a year before this guide was published. They are not a ranking, they do not describe paid versions or other tasks, and assistants have changed since. The habit that survives every model update is the same for all of them: check the claim, not the assistant.

Vendors also publish their own evaluations in system cards and model cards. Treat those as the vendor's own test: useful for comparing one vendor's versions over time, not for comparing vendors with each other.

How to fact-check AI: the 5-step check

To fact-check an AI answer, mark every checkable claim, trace each one to its original source, read across to an independent source, check the date, then keep, fix or delete it. Plan about two minutes per claim; medical, legal and financial claims need a primary source every time, however long it takes.

The seconds are our planning figures, not measurements: some claims take ten seconds, a disputed statistic can take ten minutes. They are useful for one thing, deciding before you start how long the check will take. A 1,000-word answer with 15 checkable claims needs 15 × 2 = 30 minutes.

Where to find the original for each type of claim

The original is the first publisher of the fact: the body that collected the number, the court that wrote the judgment, the person who said the words. A news article or a blog that repeats it is a pointer to the original, not a substitute.

Assistants make step 2 easier when they show sources: ChatGPT's search and deep research give cited answers [10], Gemini shows a Sources button when sources are available [13], and Anthropic asks users to review Claude's cited sources [11]. All three still leave the reading to you.

How to prevent AI hallucinations when you prompt

You cannot prevent AI hallucinations completely, but you can lower the odds and make them easier to spot: give the assistant the source material, ask for a quote and a link for every claim, make "I don't know" an acceptable answer, and switch on search for anything current.

Grounding narrows what the model can get wrong but does not remove errors: the EBU/BBC study that asked assistants to use named broadcasters' sources where possible still found significant sourcing problems in 31% of answers [1]. Prompting changes how many claims fail the check; it does not replace the check.

Are AI hallucinations getting worse?

The evidence does not show one direction: measured error rates have fallen on some tests, while documented harms are rising as more people use AI. Both statements can be true at once, so quote the measure with the number.

• Better on one comparison: in the BBC's own answers, the share with any significant issue fell from 51% to 37% between its first study and the 2025 round [1].
• Still far from solved, on Google's test: every model on the FACTS suite scored below 70% overall (December 2025) [4].
• Still far from solved, on the AI Index benchmark: hallucination rates ranged from 22% to 94% across 26 top models (2026) [3].
• More documented incidents: AI incidents of all kinds recorded in the AI Incident Database rose from 233 in 2024 to 362 in 2025, a rise of 55.4% [3].
• More court decisions: decisions involving hallucinated material reached 2,202 in Charlotin's database by 19 October 2026 [5].

Rising counts of incidents and court decisions may partly reflect more use and more reporting, not only more errors per answer. Neither source measures a rate per answer, so they cannot tell you whether a given assistant got worse.

Teaching it: a 10-minute AI-literacy lesson

A practical way to teach AI hallucinations is to let students catch one: give them a short AI answer with five checkable claims, and have them run the 5-step check in pairs. Ten minutes is enough for one round and a debrief.

The debrief is where the lesson lands: compare the class vote with the result, and ask why the answer read so well. For a full class presentation on the topic, our artificial intelligence presentation outline has a "what AI gets wrong" slide,.

Worked example: checking a 10-slide market update

A worked check shows what the 5-step method costs and what it catches. The example is fictional: Tomasz Nowak, an analyst at Brightwater Analytics, drafts a 10-slide market update with an AI assistant and checks it before it goes to a client.

Step 1, mark. Tomasz highlights 24 checkable claims across the 10 slides: 11 numbers, 5 names, 4 dates, 2 quotes and 2 citations (11 + 5 + 4 + 2 + 2 = 24). Planned time: 24 × 2 minutes = 48 minutes.

Steps 2 to 5, results.

• Deleted, 3 claims: both citations pointed to reports whose titles returned no result on the publishers' sites, and one quote could not be found in any transcript. 3 ÷ 24 = 12.5% of the claims.
• Updated, 1 claim: a market-size figure was real but two years old; the publisher had released a newer edition. 1 ÷ 24 = 4.2%.
• Confirmed, 20 claims: each now carries a source line with publisher and year. 20 ÷ 24 = 83.3%.
• Check: 3 + 1 + 20 = 24 claims, and 12.5% + 4.2% + 83.3% = 100.0%.

In Tomasz's check the citations and the quote failed while names and dates held, so next month he opens every citation first.

For scale only: if 45% of answers about the news have a significant issue, as in the EBU/BBC study, 20 such answers would contain about 0.45 × 20 = 9 that need fixing [1]. That rate was measured on news questions in 2025 and is not a forecast for any one tool or task; it is a reason to budget time for the check.

Tomasz then turns the method into a 9-slide training deck for his team. Its outline is Prompt 3 below.

Here are two slides from this article's worked example, as NOXI built them, unedited.

For presentations: check every figure before you present

A presentation multiplies a hallucination: one wrong number on a slide reaches every person in the room and every person the deck is forwarded to. Before you present an AI-assisted deck, run the 5-step check on every figure, name, quote and citation, and put a source line with the publisher and year under each number.

1. List the claims slide by slide. Count them before you start; at two minutes each, 20 claims take 40 minutes.
2. Check charts at the data, not the picture. Open the chart's numbers and compare each value with the source table; see our guide to how to present data for chart choices.
3. Keep the year with the number. "45% (EBU/BBC, 2025)" survives being copied into another deck; "45%" does not.
4. Show uncertainty honestly. If a figure is an estimate or a range, say so on the slide; our guide to presenting forecast uncertainty shows how.
5. Separate your data from published data. A survey you ran and a national statistic are two sources; label them separately, as in our survey results presentation guide.

For research talks, the same rule applies to every paper you cite: our journal club presentation and research proposal presentation guides show where citations sit on the slide. For AI tools that research a deck and put sources on its slides, see AI presentations with research and citations.

3 prompts for checkable AI answers

These three prompts turn the method into text you can paste: the first two work in any AI assistant, the third builds the training deck from the worked example.

Build a sourced training deck in NOXI

You can build the training deck from a prompt in NOXI and then run the 5-step check on it, slide by slide. The flow below uses one Home action, Create from scratch, from the first step to the download.

Frequently asked questions

What are AI hallucinations?

AI hallucinations are answers from a generative AI system that sound confident but are false, unsupported or invented, such as a wrong figure, an outdated fact, a misquoted person or a citation to a source that does not exist. NIST calls them confabulations. They come from how language models generate text: by predicting likely words rather than looking facts up.

Why does AI hallucinate?

AI hallucinates because language models predict the most likely next words, not verified facts, and know nothing after their training cutoff unless they search. An OpenAI research paper from September 2025 adds that most tests score answers right or wrong with no credit for "I don't know", so guessing is rewarded during training and evaluation.

How often does ChatGPT hallucinate?

There is no single rate; it depends on the version, tools and task. In the EBU/BBC study of free assistants answering news questions in May and June 2025, 36% of ChatGPT's answers had at least one significant issue, and 18% to 22% of every assistant's answers had significant accuracy problems. Other tasks give different figures.

Can AI make up sources?

Yes. AI can invent references with plausible authors, titles and years, and it can cite real pages that do not contain the claim. In one English High Court case decided on 6 June 2025, 18 of 45 citations put before the court did not exist. Open every link and search every title before you use it.

How do I fact-check an AI answer?

Use five steps. Mark every number, name, date, quote and citation. Trace each to its source and open the link yourself. Read across to the original publisher in a new tab. Check the date. Then keep the claim with its source, fix it from the original, or delete it. Plan about two minutes per claim.

How accurate is ChatGPT?

ChatGPT's accuracy varies by version, task and whether search is on, and OpenAI's own help centre says it can produce incorrect or misleading outputs and should be used as a first draft, not a final source. Independent studies measure one task at one time, so treat any single percentage as that study's result only.

Are AI hallucinations getting worse?

Not in every measure. In the BBC's own answers, significant issues fell from 51% to 37% between its first study and 2025. Yet every model on Google DeepMind's FACTS suite scored below 70% in December 2025. Separately, court decisions involving hallucinated material reached 2,202 by 19 October 2026 as AI use grows.

What happens when lawyers cite AI-invented cases?

Courts hold the lawyer responsible for every citation filed. The English High Court said on 6 June 2025 that lawyers must check AI research against authoritative sources and referred a lawyer to the regulator. On 3 June 2026 the US Ninth Circuit suspended two attorneys from practice before it for six months and imposed monetary sanctions.

How can I reduce hallucinations in my prompts?

Give the assistant the source document and ask it to answer only from that document. Ask for a quote and a link for every claim. Tell it that "I don't know" is acceptable, for example by asking it to answer only above 90% confidence. Switch on search for current facts, and ask neutral rather than leading questions.

How do I check the numbers in an AI-made presentation?

List every number slide by slide, open each source line and find the figure on the linked page, then confirm the year and whether a newer edition exists. Check charts against their data table, not the picture. Put the publisher and year under every number, and delete any figure whose source you cannot open.

Sources & methodology

How this guide was made: every rate, study and case was taken from the primary source, opened on 22 October 2026: the EBU/BBC report and release, the Stanford HAI 2026 AI Index chapter, Google DeepMind's FACTS post, Vectara's leaderboard, Charlotin's database, the English and US judgments, NIST AI 600-1, the OpenAI paper and the vendors' help pages. Each figure is quoted with the measure, sample and date its source gives; figures from different studies are never combined. The 5-step check, its timings, the lesson plan and the prompts are editorial recommendations. Tomasz Nowak and Brightwater Analytics are fictional; every calculation was checked twice. All sources were checked on 22 October 2026.

• [1] 22 public service media organisations in 18 countries and 14 languages; free/consumer versions of ChatGPT, Copilot, Gemini and Perplexity; responses generated late May and early June 2025; 45% with at least one significant issue, 81% including some issues; sourcing 31%, accuracy 20%, context 14%; direct quotes 12%; per-assistant significant issues (ChatGPT 36%, Copilot 37%, Gemini 76%, Perplexity 30%); Gemini sourcing 72%; accuracy 18% to 22% for all assistants; BBC-to-BBC comparison 51% to 37% (362 and 237 responses); examples on the Pope, bird flu and vapes — EBU and BBC, News Integrity in AI Assistants report (October 2025).
• [2] More than 3,000 responses; 45%, 31% and 20% headline figures; study led by the BBC and coordinated by the EBU (22 October 2025) — EBU, Largest study of its kind shows AI assistants misrepresent news content 45% of the time.
• [3] Hallucination rates across 26 top models from 22% to 94% on a new accuracy benchmark (AA-Omniscience in the chart caption); performance collapses when a false statement is presented as the user's belief; AI incidents 362 in 2025, up from 233 in 2024 — Stanford HAI, The 2026 AI Index Report, chapter 3: Responsible AI.
• [4] Four benchmarks (parametric, search, multimodal, grounding v2), 3,513 examples, FACTS Score as average accuracy, all evaluated models below 70% overall, managed with Kaggle (9 December 2025) — Google DeepMind, FACTS Benchmark Suite.
• [5] 2,202 cases identified; scope limited to decisions where a court or tribunal found or implied reliance on hallucinated content; updated 19 October 2026 — Damien Charlotin, AI Hallucination Cases database.
• [6] [2025] EWHC 1383 (Admin), Divisional Court, 6 June 2025: paragraphs 6 to 8 on generative AI and the duty to check against authoritative sources; five fake cases in Ayinde (paragraph 43); 45 citations, 18 nonexistent, in Al-Haroun (paragraph 74); decision not to initiate contempt proceedings and referral (paragraphs 68 to 70) — Courts and Tribunals Judiciary, Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank.
• [7] No. 24-4790, order filed 3 June 2026: sanctions for nonexistent cases, misattributed quotations and lack of candour; rules violated at the point of signing and filing; six-month suspension and monetary sanctions; "read everything cited in a court filing" — US Court of Appeals for the Ninth Circuit, Lnu v. Blanche, order.
• [8] Definition of confabulation (section 2 and section 2.2), its causes, the confident-response risk and the footnote on anthropomorphising terms (July 2024) — NIST AI 600-1, Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile.
• [9] Kalai, Nachum, Vempala and Zhang, 4 September 2025: hallucinations as classification errors; the 20% birthday lower bound; binary grading rewards guessing; the three-wrong-birthdays example; explicit confidence targets with a penalty of t/(1 − t) — OpenAI, Why Language Models Hallucinate (paper).
• [10] Hallucination examples, knowledge cutoff, "Confidence isn't reliability", search and deep research for cited answers, first draft not a final source (opened 22 October 2026) — OpenAI Help Center, Does ChatGPT tell the truth?.
• [11] Claude can occasionally produce incorrect or misleading responses; not a singular source of truth; review cited sources with web search (17 March 2026) — Claude Help Center, Claude is providing incorrect or misleading responses.
• [12] Gemini Apps may produce inaccurate information; not professional advice; double-check responses about people; Gemini can hallucinate, including about how it works — Gemini Apps Help, Learn about responses from Gemini Apps.
• [13] The Sources button at the bottom of a response or in line, when sources are available — Gemini Apps Help, View related sources from Gemini Apps.
• [14] Hallucination rates when summarising a document, 1.8% to 24.2%, scored with Vectara's HHEM-2.3 model; dataset of over 7,700 articles; last updated 16 October 2026; run by Vectara — Vectara, Hallucination Leaderboard (GitHub).

Build a training deck on checking AI answers in NOXI, free: describe it from a prompt, review the outline in Plan mode, open every source line, then present full-screen or download an editable PowerPoint.