AI Hallucination Examples: 5 Real Cases, With Sources
Most lists of AI hallucination examples repeat the same screenshots with no source. This one has five real cases, and each links to the primary record: a tribunal decision, a federal court order, Google's own blog post, and answers we recorded from Claude on 15 October 2026.
Three of the five are about a business: what it owns, who runs it, and what its refund policy says. That is where an AI error costs money.
What is an AI hallucination?
OpenAI defines hallucinations as "plausible but false statements generated by language models". The word covers three different failures:
- Invented: nothing true sits behind it. A court case that does not exist.
- Outdated: it was true once. The company sold the product, or the CEO changed.
- Misread: a source exists, but the AI used the wrong one or read it wrong.
5 real AI hallucination examples
1. Claude says Under Armour owns MyFitnessPal (2026)
Our first example of AI hallucination comes from our AI visibility audit of Nike, Adidas, Puma and Under Armour. On 15 October 2026 we asked Claude (claude-sonnet-5, web access off) "Under Armour vs Nike". It wrote:
"Connected fitness (MyFitnessPal, MapMyRun) – UA owns popular fitness-tracking apps, giving it a data/engagement edge Nike doesn't fully match"
Under Armour owns neither app. It completed the sale of MyFitnessPal to Francisco Partners on 18 December 2020 (SEC filing), and Outside Inc. announced its acquisition of MapMyFitness, which includes MapMyRun, on 30 August 2024.
In another answer, the same model wrote that connected fitness was "largely divested — MyFitnessPal was sold in 2020". The model had the right fact. Which version came out depended on the question. Type: outdated.
2. Claude gives PUMA's CEO two different start years (2026)
Asked "who is the CEO of Puma", Claude named Arthur Hoeld, correctly, and said he took over "in 2024". It then hedged: "I should flag some uncertainty here—executive positions change frequently". Asked for Puma's history, it gave a different year: "Arthur Hoeld took over as CEO in 2023".
PUMA's own management page lists him as a member of the Management Board since July 2025. Right name, two wrong years. Type: an invented detail on a real change.
OpenAI's researchers saw the same thing: asked for one author's birthday, a widely used chatbot "gave three different dates, likewise all wrong."
3. Air Canada's chatbot misstates its own refund policy (2024)
In November 2022, after their grandmother died, Jake Moffatt asked the chatbot on Air Canada's website about bereavement fares. It told them:
"If you need to travel immediately or have already travelled and would like to submit your ticket for a reduced bereavement rate, kindly do so within 90 days of the date your ticket was issued by completing our Ticket Refund Application form."
Air Canada's own bereavement page said the policy did not apply once travel was completed. The airline argued it could not be held liable for its chatbot. The tribunal called that "a remarkable submission":
"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."
In Moffatt v. Air Canada, 2024 BCCRT 149 (14 February 2024), the airline was ordered to pay $812.02: $650.88 in damages, plus interest and fees. One caution: Air Canada "did not provide any information about the nature of its chatbot", so the decision does not show a generative AI model wrote the answer. It does show that a company answers for what its chatbot tells customers.
4. ChatGPT invents six court cases (2023)
In a passenger's lawsuit against the airline Avianca, his lawyers filed a brief citing cases that did not exist. The court's opinion and order on sanctions (S.D.N.Y., 22 June 2023) says they "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 "decisions" were ChatGPT's work. Asked "Is Varghese a real case", ChatGPT answered, per the order, that it had supplied "real" authorities that could be found on Westlaw and LexisNexis. The court imposed a $5,000 penalty. It did not sanction the use of AI as such: "there is nothing inherently improper about using a reliable artificial intelligence tool for assistance." Type: invented, in full.
5. Google AI Overviews answers "How many rocks should I eat?" (2024)
In "AI Overviews: About last week" (30 May 2024), Google conceded that "some odd, inaccurate or unhelpful AI Overviews certainly did show up." Satirical content about eating rocks had been republished on a geological software provider's website, and "an AI Overview appeared that faithfully linked to one of the only websites that tackled the question." Forum content led to "less-than-helpful advice, like using glue to get cheese to stick to pizza."
Google says AI Overviews "generally don't 'hallucinate'" the way other LLM products might; when they go wrong, it is usually from misreading queries or language on the web. For a brand, that is the warning: an AI that searches can faithfully repeat a wrong page about you. Type: misread.
AI hallucination cases at a glance
| Case | What the AI said | What was true | Type |
|---|---|---|---|
| Claude on Under Armour (2026) | UA owns MyFitnessPal and MapMyRun | Sold in 2020 and 2024 | Outdated |
| Claude on PUMA (2026) | CEO since 2023, or since 2024 | CEO since July 2025 | Invented detail |
| Air Canada chatbot (ruled 2024) | Claim the fare within 90 days, even after travel | Not after travel | Wrong policy (cause not stated) |
| ChatGPT in Mata v. Avianca (2023) | Six court opinions | None existed | Invented |
| Google AI Overviews (2024) | Answers built on satire and forum advice | Satire and a forum post | Misread |
Invented vs outdated: why the difference matters for a brand
Invented errors are hard to predict. OpenAI's researchers write that "arbitrary low-frequency facts, like a pet's birthday, cannot be predicted from patterns alone and hence lead to hallucinations." A CEO's start year is that kind of fact.
Outdated errors are predictable. They sit on facts with a before and an after: an owner, a CEO, a price, a product line you sold or dropped.
Check the facts that changed first. All three errors in our audit sat there, while all 16 answers to "who is the CEO" named the right person.
A model answering from memory, as Claude did here, keeps its answer until the vendor ships a new model; nothing you publish this quarter reaches it. An AI that searches reads the live web, so the current fact has to sit on a page it can reach. For the steps, see how to find and fix AI hallucinations about your brand.
Check your own chatbot too. Moffatt v. Air Canada is one small-claims decision in one Canadian province. Its reasoning is still plain: the chatbot was part of the website, so the company answered for it.
How to check if AI is hallucinating about your brand
- List the facts that changed in the last five years. Owner, CEO, prices, products sold or dropped, headquarters.
- Ask each AI platform about each fact, in more than one wording. A direct question and a comparison with a rival: Claude got the MyFitnessPal sale right in one and wrong in the other.
- Check every claim against a primary source. A filing, a press release, your own page. Never another AI answer.
- Record the date, the platform, the model and whether it searched the web. Without them, you can't compare next month's check.
- Ask again after each model update. Memory answers change only when the model does.
We found all three of our errors by reading the answers by hand. 99Visibility runs audits across five AI platforms (ChatGPT, Gemini, Perplexity, Google AI Overviews and Claude) and checks what the answers say against the brand facts you enter. Your first audit is free, with no credit card, and covers ChatGPT and Google AI Overviews. See the plans for all five platforms. Inside the dashboard, Agent V reads your audits and your site and drafts the fix, such as page copy for you to paste. It never edits your site itself.
FAQ
Which AI hallucinates the most?
Our data can't say. All three confirmed errors in our 67 answers came from Claude, but three errors are too few to rank platforms, and the platforms answered under different conditions: ChatGPT and Claude from memory, Perplexity with live search, Gemini cut short by a token limit on our side. Before trusting any leaderboard, check what it measured and on which model version.
Why does ChatGPT hallucinate?
OpenAI says language models hallucinate "because standard training and evaluation procedures reward guessing over acknowledging uncertainty." In its words, "the guessing model ends up looking better on scoreboards than a careful model that admits uncertainty."
See what AI says about your brand. Run a free AI visibility audit: no credit card needed.
Want more of our research in Google Search and AI Mode? Add us as a preferred source.
See what AI says about your brand
Your free audit asks ChatGPT and Google AI Overviews about your brand. Paid plans add Gemini, Perplexity and Claude, and show every answer that contradicts your facts.
Start Free Audit- ChatGPTMentioned, 2nd of 5 brands
- Google AI OverviewsNot mentioned
- PerplexityMentioned and cited as source [1]
- ClaudeSays $29/mo. Your pricing page says $49/mo.
Related Articles
ChatGPT Wrong Information About My Company? How to Fix It
There is no form that edits what ChatGPT says about you. Here is what works: find where the wrong fact came from, fix the source, and use each platform's report route.
We Ran an AI Visibility Audit on Nike, Adidas, Puma and Under Armour. Here's What AI Got Wrong
We asked ChatGPT, Perplexity and Claude the questions sportswear shoppers ask, then checked the answers by hand. Nike was named first in 13 of 15 answers. Under Armour was named first in none, and Claude says it still owns MyFitnessPal.
What Does ChatGPT Say About My Brand? How to Check It
A step-by-step way to see what ChatGPT says about your brand: the prompts, search off and on, scoring against your own facts, and what a manual check cannot see.