# AI mistakes and hallucinations: how to check the answers

> Why ChatGPT, Claude, Gemini, and Copilot make things up, real cases from courts in four countries, what studies measured, and an eight-step way to check.

Yes, ChatGPT and every other AI assistant make mistakes. An **AI hallucination** is a plausible but false statement given with confidence, such as an invented date, quote, or source. OpenAI, Anthropic, Google, and Microsoft all say their tools can do this, and none promises zero errors. Courts and tribunals in the US, the UK, Canada, and Australia have already ruled on the consequences.

**Key takeaways**

1. In Moffatt v. Air Canada, a British Columbia tribunal ruled in February 2024 that the airline was responsible for its chatbot's wrong answer about bereavement fares and awarded CA$812.02.
2. A Stanford study of paid AI legal research tools, with questions asked in 2024 and published in 2025, found hallucinations in 17% to 33% of answers.
3. OpenAI researchers wrote in September 2025 that language models are trained and evaluated in ways that reward guessing over admitting uncertainty.


## Why AI gets things wrong, and how to catch it

An AI assistant can be wrong in the same calm, fluent tone it uses when it is right, which is what makes its mistakes expensive. This page is part of [getting work done with everyday AI](/guides/ai-at-work/everyday-tasks/) and explains, from the companies' own help pages, court decisions, and published studies read on November 10, 2026, why it happens and how to check. OpenAI's own help page asks you to [use ChatGPT as a first draft](https://help.openai.com/en/articles/8313428-does-chatgpt-tell-the-truth), not a final source.

The short answer to “can ChatGPT be wrong” is yes, about facts, dates, figures, quotes, and sources, and the same goes for Claude, Gemini, Copilot, and Perplexity. Checking is part of the job, not a sign of distrust.

### What a hallucination looks like

The US National Institute of Standards and Technology calls it confabulation: false content stated with confidence, more frequent in long answers and specialized subjects, sometimes with invented reasoning or citations to back it up. The companies list the forms it takes:

- wrong dates, facts, figures, or definitions;
- quotes, studies, and references that do not exist;
- confident answers to ambiguous questions;
- outdated information presented as current;
- claims about what the assistant did, such as “I sent the email,” which Anthropic says Claude can make without having any tool to do it;
- mistakes about itself: Google says Gemini can be wrong about how it works.

### Why it happens

OpenAI's researchers gave their explanation in September 2025. A language model predicts likely text, and the tests used to train and rank models score it like a multiple-choice exam, where a guess earns more than a blank. So models learn to guess. A fact that appears only once in the training data, such as a person's birth date, is hard to learn at all. Web search helps, but they add that it is no cure, and it does nothing for a failed calculation.

### When it went wrong in practice

| Case | Where and when | What happened | Outcome |
|---|---|---|---|
| *Mata v. Avianca* | Federal court in New York, June 22, 2023 | Lawyers filed court decisions invented by ChatGPT and stood by them after the judge questioned them | $5,000 penalty |
| *Harber v. HMRC* | UK tax tribunal, December 4, 2023 | A taxpayer contesting a £3,265.11 penalty cited nine tribunal decisions, all generated by AI | Appeal dismissed; the tribunal accepted she did not know |
| *Moffatt v. Air Canada* | British Columbia tribunal, February 14, 2024 | The airline's chatbot said a bereavement fare could be claimed after travel; the page it linked to said otherwise | Air Canada held responsible for all the information on its website; CA$812.02 |
| *Ayinde* and *Al-Haroun* | High Court in London, June 6, 2025 | In a £89.4 million claim, 18 of 45 cited cases did not exist; the client had done the research | Lawyers referred to their regulators |

[*Moffatt*](https://decisions.civilresolutionbc.ca/crt/crtd/en/525448/1/document.do) is the one to remember if your business has a chatbot: the tribunal rejected the idea that the chatbot was responsible for its own actions and found that a customer had no reason to double-check one part of the website against another. [*Harber*](https://caselaw.nationalarchives.gov.uk/ukftt/tc/2023/1007) shows how plausible a fake can be: the invented decisions used the names of real parties with different years and outcomes, repeated identical phrases, and used American spelling in a British case.

### What studies have measured

| Study | What was tested | Result |
|---|---|---|
| Stanford RegLab and HAI, questions asked in March and April 2024 | 202 legal questions put to paid legal research tools | Hallucinations in 17% to 33% of answers, despite claims of “hallucination-free” tools |
| Journal of Medical Internet Research, May 2024 | References for medical literature reviews from three chatbots | From 28.6% to 91.4% of references invented, depending on the chatbot |
| Tow Center, Columbia Journalism Review, March 2025 | 1,600 requests to eight AI search tools to identify a news article from an extract | More than 60% of answers incorrect overall |
| European Broadcasting Union and BBC, October 2025 | Over 3,000 answers about the news from four assistants | 45% with at least one significant issue, according to the EBU |

These tests used models from 2023 to 2025, which have changed since. The [Stanford study](https://arxiv.org/abs/2405.20362) says the same of itself: it is a snapshot. Microsoft describes its own accuracy evaluations for Copilot without publishing an error rate.

### Eight steps to check an answer

This is our method, built from the companies' advice and the courts' findings:

1. **Mark what must be checked:** numbers, dates, names, quotes, references, anything about a real person, and any claim that the assistant did something.
2. **Give it your documents** and ask it to use only those. Tools grounded in your files, such as Gemini Notebook, reduce invention without removing it.
3. **Allow “I don't know”** in your prompt, so a guess is not the only way out.
4. **Ask for a source for each claim,** and drop any claim that has none, a technique Anthropic documents.
5. **Open the sources:** check that the link works, that the page says what the answer claims, and when it was written. A citation can be invented or point to a page that says something else.
6. **Redo the arithmetic** in a spreadsheet, or use ChatGPT's data analysis tool, which runs the calculation as code.
7. **Ask again and compare.** Two different answers to the same factual question are a warning sign.
8. **Find the original** of every law, decision, or study in an official database, and have a second person review anything that leaves the business.

The [UK judiciary's guidance on AI](https://www.judiciary.uk/wp-content/uploads/2025/10/Artificial-Intelligence-AI-Guidance-for-Judicial-Office-Holders-2.pdf) adds warning signs that apply to any document: references nobody recognizes, spelling from another country, very persuasive text that is wrong on substance, and leftover phrases such as “as an AI language model.”

### AI errors in a business report

Microsoft says that even when Copilot summarizes your company's own documents, it can add information that is not in them. In October 2025, Australia's employment department said Deloitte had confirmed that some footnotes and references in a review it delivered were incorrect, and published a corrected version; press reports linked the errors to generative AI, which the department's pages we read do not mention. In an Australian Federal Court case decided in 2025, the court found that the central error behind false citations was the failure to check the output, together with the lack of supervision of a junior lawyer. The lesson for a small team is the same: someone other than the author checks the sources before a report goes out.

### Prompts that leave room for doubt

Steps 2 to 4 start in the prompt. See [how to write a prompt that gets useful work](/guides/ai-at-work/everyday-tasks/write-better-prompts/), with a line that asks the assistant to show its work.

### Summaries that add things

Summaries fail in their own ways, from skipped pages to invented sources. Our page on [checking an AI summary](/guides/ai-at-work/everyday-tasks/summarize-documents/) covers them.