Reading time: 13 minutes
AI hallucination is when a model like ChatGPT makes up or distorts information and presents it as fact. The types of AI hallucinations range from a wrong number to a source that does not exist. The hallucination rate varies by model and test, but it never reaches zero. So Google and Air Canada have already paid for chatbot errors.
In short
- According to OpenAI, a hallucination is a plausible but false statement. That is why a wrong answer sounds as confident as a right one.
- A research review splits the types of AI hallucinations into two groups: errors against the facts and errors against the request or the text provided.
- The hallucination rate depends on the test. When summarizing texts, 108 models got it wrong 1.8% to 24.2% of the time, according to Vectara.
- Less-cited brands face more risk, because the model learns through repetition. Web search helps, but it does not end the errors: in an EBU study with the BBC, 45% of answers about the news had at least one significant issue.
What is AI hallucination?
AI hallucination is a false or baseless answer that a model presents as true. It can swap a number, invent a name or cite a source that does not exist. After all, the model writes the most likely text, not checked text. That is why the error often comes in the same confident tone as a correct answer.
In its Help Center, OpenAI gives three examples of hallucination. The list includes incorrect facts or dates, made-up references and overconfident answers to ambiguous questions. In addition, the company warns that ChatGPT can sound confident even when it is wrong. In another text, from 2025, OpenAI sums it up: hallucinations are plausible but false statements.
The Cambridge Dictionary chose “hallucinate” as its Word of the Year for 2023. Its entry then gained a new meaning: when an AI hallucinates, it produces false information. But not every AI error comes from a hallucination. An old price, for example, may come from the training cutoff, while a wrong address may come from a wrong web page. The causes of each error are in our guide on why ChatGPT makes mistakes.
What are the types of AI hallucinations?
The types of AI hallucinations fall into two broad groups. In the first, the answer contradicts real-world facts. In the second, however, it contradicts the user’s request, the text the user provided or its own logic. This split of the types of AI hallucinations comes from a review accepted by ACM Transactions on Information Systems.
In the review, Lei Huang and coauthors call the first group factuality hallucination. They call the second faithfulness hallucination. According to the authors, contradiction is the most frequent type of factuality hallucination. The table shows each type of AI hallucination, with an example of how it can show up for a company.
| Type | What happens | Example |
|---|---|---|
| Factual contradiction | Real fact with a wrong detail | Wrong opening hours or price |
| Factual fabrication | Fact with no basis in the real world | Made-up partner, award or lawsuit |
| Instruction inconsistency | Does something other than what was asked | Asks to translate a question and gets the answer |
| Context inconsistency | Contradicts the text provided | Website summary with a deadline the site does not mention |
| Logical inconsistency | Contradicts itself | Sum of the plans with the wrong total |
For a brand, the types of AI hallucinations have different effects. According to Wolf River Electric’s lawsuit, described below, the AI tied the company to a real lawsuit against others: a factual contradiction. Context inconsistency, on the other hand, shows up when the AI reads the right source and summarizes it wrong, as in the Apple case.
What is the hallucination rate of the leading AI models?
The hallucination rate depends on the test. When it summarizes a text it was given, a typical model gets it wrong close to 10% of the time. However, when it answers hard questions from memory, many models would rather guess than admit they do not know. In the two tests below, no AI model reaches zero, and the numbers change with each version.
In the Vectara ranking, updated in September 2026, 108 models summarized more than 7,700 texts. The rule was to use only what was in each text. Then Vectara’s own detector flagged the summaries that included something not in the text. In this test, the hallucination rate ranged from 1.8% to 24.2%, and the median, by our count, was 9.55%. GPT-6 Astra, OpenAI’s main model, got 8.7% of its summaries wrong. The preview version of Google’s Gemini 3.1 Pro got 10.4% wrong, and Anthropic’s Claude Opus 4.7, 12%. But Vectara sells that detector.
Which AI hallucinates the least?
In AA-Omniscience, from Artificial Analysis, Google’s Gemini 4 Argon had the lowest hallucination rate among the 32 featured models. The test asks 6,000 hard questions, without search, and tells the model to answer only when it is confident. The rate counts only the questions the model did not get right. Among those, it shows how often the model guessed wrong instead of admitting it did not know.
| Model | Company | Hallucination rate |
|---|---|---|
| Gemini 4 Argon | 15.1% | |
| GPT-6 Astra | OpenAI | 44.8% |
| Claude Opus 5.5 | Anthropic | 58.6% |
| GPT-6 Luna | OpenAI | 76.7% |
| DeepSeek V4.1 Flash | DeepSeek | 96.5% |
The table shows the top model from Google, OpenAI and Anthropic, plus two cheaper models, all released in September 2026. When the ranking shows more than one version of the same model, the table uses the one with the best score on the test. The two OpenAI models show that one company can have very different rates. But a low rate does not mean more correct answers: Gemini 4 Argon, first released to a restricted group, left more questions unanswered. Overall, it got 49.9% of the questions right, compared with 66.2% for Claude Opus 5.5.
For brand facts, Searchable found that 8.1% to 11.5% were false in the three AI engines that brands rated most. The company, however, sells AI brand monitoring.
Why does AI hallucinate more about less-cited brands?
Because the model learns through repetition. A fact that appears in many texts becomes a pattern, but a fact that appears only once has no support. So when someone asks about a small brand’s partner or address, the model tends to fill the gap with something plausible.
In a 2025 paper, researchers from OpenAI and Georgia Tech put numbers on this effect. If 20% of birthday facts appear only once in the training data, the model should hallucinate on at least 20% of them. This holds for the base model, before the final tuning. In other words, the hallucination rate follows how rare a fact is. And small-company data, such as partners and addresses, tends to be rare in texts.
A study presented at the ICML 2023 conference reached the same point. In it, the BLOOM model got 25% of fact questions right when about 10 supporting documents were in its training data. With about 10,000 documents, however, it got more than 55% right.
The effect shows up even with search. In an experiment published in 2024, ChatGPT, connected to the internet, suggested 20 digital marketing agencies in each of three countries. In Germany, 90% of the suggestions were correct; in Poland, 75%; and in Latvia, the smallest market, only 35%. That is why a brand’s defense is to multiply the independent, consistent sources about it. See which sources AI cites most and how to earn a place in them.
Does web search reduce AI hallucination?
It does, but it does not end it. With search, the hallucination rate drops, because the model answers based on pages it has just read. This helps most with rare facts, such as those about small brands. However, search does not prevent all types of AI hallucinations, because the AI can still misread the page. In addition, it can repeat a wrong source.
In an OpenAI test from March 2025, GPT-4o got 38% of the factual questions in SimpleQA right without search. The search version of GPT-4o got 90% right. Likewise, the ICML study showed that looking up documents at answer time makes a model rely less on its training for rare facts.
Even so, search makes mistakes. In a 2025 study coordinated by the EBU and led by the BBC, journalists evaluated more than 3,000 answers about the news. The answers came from ChatGPT, Copilot, Gemini and Perplexity. Overall, 45% had at least one significant issue, and 20% had major accuracy issues, such as hallucinated details. Gemini had issues in 76% of its answers, mostly because of its sourcing.
Google described the same limit for AI Overviews in a May 2024 statement. According to the company, they generally do not make things up the way other AI products might. When they get it wrong, the cause is usually a misread query, a nuance of language on the web or a lack of good information. That is how one summary suggested glue to make cheese stick to pizza, based on forum content. The full path of an answer, from training to search, is in our guide on how ChatGPT works.
What are some real examples of AI hallucination involving companies?
There are cases in courts, in the press and in companies’ own statements. They show three risks: being liable for what a bot made up, losing market value and seeing the brand tied to a false fact. The table sums up six cases from 2023 to 2025 linked to AI hallucination, with caveats right below it.
| Company | The error | What happened |
|---|---|---|
| Google (2023) | False claim about the James Webb telescope in a Bard ad | Alphabet lost $100 billion in market value |
| Air Canada (2024) | Refund rule that did not exist, given by the chatbot | Tribunal ordered it to pay 812 Canadian dollars |
| Apple (2024) | False news in a summary of a BBC alert | Apple suspended its news summaries |
| Cursor (2025) | One-device-per-subscription limit, made up by the support bot | Company apologized and refunded the customer |
| Wolf River Electric (2025) | Attorney general lawsuit that did not exist, in an AI Overview | Company is suing Google for defamation, and Google denies it |
| Deloitte (2025) | Nonexistent references in a report made with AI help | Agreed to repay the final installment to Australia’s government |
Sources: NPR (Google), Civil Resolution Tribunal (Air Canada), BBC (Apple), The Register (Cursor), Star Tribune (Wolf River) and The Guardian (Deloitte). Caveats: the Air Canada decision does not detail the chatbot’s technology, and Deloitte admitted using AI but did not attribute the errors to it.
Is a company liable for what its chatbot says?
At least in the Air Canada case, yes. Before the tribunal, the airline argued it was not liable for the chatbot’s information, as if the chatbot were a separate entity. In February 2024, British Columbia’s Civil Resolution Tribunal rejected that argument, in a decision that became a reference:
“It should be obvious to Air Canada that it is responsible for all the information on its website. It makes no difference whether the information comes from a static page or a chatbot.”
Christopher C. Rivers, Civil Resolution Tribunal member (2024)
The Wolf River Electric case shows the risk even for companies that do not use AI. According to the lawsuit, described by the Star Tribune, an AI Overview said Minnesota’s attorney general was suing the solar company. In fact, that lawsuit targeted four other companies, and none of the four sources the AI cited supported the claim. The company says it lost a $150,000 contract. However, these are its allegations in the lawsuit, and Google denies them in its response.
If AI made something up about your company, start with the source of the error and see how to report a ChatGPT error. Then repeat the same questions for a few weeks, as our guide on how to measure brand mentions and citations explains.
What we found analyzing 13 Portuguese-language pages on AI hallucination
We analyzed the 13 Portuguese-language pages that came up in 6 searches on the topic. The searches included “o que é alucinação de IA” (what is AI hallucination) and “casos de alucinação de IA” (cases of AI hallucination). Eleven pages come from companies that sell AI products or courses, and only two come from news outlets.
Six list types of AI hallucinations, but only two separate the two main groups: errors against the facts and errors against the text provided. Only one gives a hallucination rate by model, and for 2025 models. Five cite Google’s Bard or the Air Canada chatbot, cases from 2023 and 2024.
However, none describes a case like Wolf River’s, in which AI made up a fact about a company. Nine say that connecting AI to sources reduces hallucination, but only one gives error data for assistants that search the web. Finally, two say that obscure or niche topics cause more hallucination, but none links this to brands.
Frequently asked questions
Not always. According to OpenAI, ChatGPT can give incorrect or misleading answers and sound confident even when it is wrong. That is why the company itself advises checking important information against reliable sources and opening the links in the answer when search is on.
Not completely. According to OpenAI, no model will reach 100% accuracy, because some questions have no answer. However, AI hallucination is not inevitable, since a model can abstain and say it does not know. In addition, web search lowers the hallucination rate, and human review checks what goes out to the public.
In a hallucination, the AI makes up or distorts a fact. With outdated information, it repeats a fact that used to be true, such as an old price. This happens because the model learns from texts up to a cutoff date. So it does not know what came later, unless it uses search. To the reader, both errors look the same.
Want to know how your brand shows up in AI answers today?
NotHuman is opening spots for agencies and brands to track how they appear in AI answers. Join the waitlist and we’ll let you know when a spot opens up for you.
Main sources
- OpenAI, Does ChatGPT tell the truth? (2026).
- Huang and coauthors, A Survey on Hallucination in Large Language Models (2024).
- Vectara, Hallucination Leaderboard, a ranking of hallucination in summaries (2026).
- Artificial Analysis, AA-Omniscience, a knowledge and hallucination test (2026).
- Kalai and coauthors, Why Language Models Hallucinate (2025).
- Prohorovs, Dalal and Radulescu, study on ChatGPT in the search for business partners (2024).
- BBC, EBU study with the BBC on AI assistants and the news (2025).
- Civil Resolution Tribunal, Moffatt v. Air Canada decision (2024).
