What Is AI Hallucination? Real Examples and How to Spot Them
Key Takeaways+
- An AI hallucination is false information an AI presents as fact, often with full confidence.
- AI predicts likely words. Unless it’s connected to your real data, it doesn’t check facts.
- Air Canada, Deloitte and a New York law firm all paid for errors their AI tools made up.
- Warning signs include missing sources, links that don’t open and policies that aren’t on your site.
- A human check before anything reaches a customer, a client or the public stops most of the damage.
An AI hallucination is when an AI tool states something false as if it were fact. It might invent a policy, a source, a number or a court case. The answer sounds confident, so people believe it. That risk grows as teams build AI into everyday workflows, where one wrong answer can reach a customer before anyone checks it. Take one AI hallucination example: A chatbot promises a refund your policy doesn’t allow. The customer believes it, and your business pays.
The business pays for the mistake, not the AI. In 2024, a Canadian tribunal ruled that Air Canada had to pay for a refund policy its chatbot made up. At Dscvred Studio, we build AI automation for finance, healthcare and B2B SaaS teams. We place a human check wherever a wrong answer would cost money or trust.
What Does AI Hallucination Mean?
AI hallucination means a generative AI model gives an answer that is false or unsupported. It then presents that answer as true. The term covers large language models like ChatGPT, Claude and Gemini, plus AI tools built on them.
Picture a support bot that tells a customer they can get a refund within 90 days. Your policy says 30. The bot didn’t lie on purpose. It wrote the answer that sounded most likely.
AI hallucinations in business usually take one of these forms:
- Made-up facts: A wrong date, statistic or product detail
- Fake sources: Citations, reports or links that don’t exist
- Invented policies: Refund rules, prices or terms your company never set
- Wrong numbers: Totals, rates or figures with no source behind them
- False actions: An AI agent says it completed a task it never did
Why Does AI Hallucinate?
AI hallucinates because it predicts the next likely word rather than looking facts up. A large language model learns patterns from training data. When it lacks a fact, it fills the gap with something that fits the pattern.
Two things make this worse:
- No grounding: If the AI isn’t connected to your approved documents, it answers from memory.
- Rewarded guessing:OpenAI’s 2025 research found that standard tests reward a confident guess over “I don’t know”. So models learn to guess.
Connecting AI to verified sources, often called retrieval-augmented generation (RAG), reduces hallucinations. It doesn’t remove them.
7 AI Hallucination Examples From Real Businesses
Each AI hallucination example below is a real, reported case. For each one, we note what it cost and the check that would have caught it.
1. KPMG and EY pull reports with fake citations (2026)
In May 2026, EY removed a report on loyalty programmes. Researchers at GPTZero found made-up data and a cited McKinsey study that doesn’t exist. Weeks later, KPMG pulled a report on agentic AI. GPTZero found that 40 of its 45 citation titles appeared fake.
What it cost: Both firms removed the reports and opened reviews. The false figures had already been repeated by trade press.
The check that catches it: Open every source and confirm it says what the report claims, before anything is published.
2. Air Canada’s chatbot invents a refund policy (2024)
A customer asked Air Canada’s website chatbot about bereavement fares. It told him he could claim the discount after travel. The airline’s real policy didn’t allow that.
What it cost: A tribunal ordered Air Canada to pay just over CA$650 in damages, plus interest and fees. It also rejected the airline’s argument that the chatbot was responsible for its own words.
The check that catches it: Answer only from approved policy pages. Send refund questions to a person.
3. Lawyers cite fake court cases from ChatGPT (2023)
In Mata v. Avianca, two New York lawyers filed a brief with cases ChatGPT had made up. When asked, ChatGPT even said the cases were real.
What it cost: A federal judge fined the lawyers and their firm $5,000.
The check that catches it: Look up every citation in a legal database before filing.
4. Deloitte’s government report cites sources that don’t exist (2025)
Deloitte Australia delivered an AU$440,000 report to a federal department. Readers found fake academic references and a made-up quote from a court judgment. Deloitte later said it had used Azure OpenAI GPT-4o in the work.
What it cost: Deloitte repaid the final payment on the contract and reissued the report.
The check that catches it: An expert checks every reference before a client deliverable goes out.
5. Cursor’s support bot invents a one-device rule (2025)
Users of the coding tool Cursor were being logged out when they switched devices. The AI support agent said this was a new policy. It wasn’t. A security update had caused a bug.
What it cost: Users threatened to cancel, and the story spread on Reddit and Hacker News. Cursor apologized and refunded affected users.
The check that catches it: Block any reply that states a policy not found in your help docs. Label AI replies, which Cursor now does.
6. A newspaper prints a reading list of fake books (2025)
The Chicago Sun-Times ran a summer reading list in a licensed insert. Most of the 15 books on it didn’t exist, though real authors’ names were attached. The freelance writer later said he used AI and didn’t check it.
What it cost: Public criticism and a statement from the paper distancing itself from the content.
The check that catches it: An editor verifies every title and name before publishing.
7. Google Bard’s launch demo gets a space fact wrong (2023)
In a promotional demo, Google’s Bard made a claim about the James Webb Space Telescope. It said the telescope took the first picture of a planet outside our solar system. Astronomers pointed out that was wrong.
What it cost: Alphabet lost about $100 billion in market value the day the error was reported.
The check that catches it: Fact-check every public claim before launch, however small.
How Can You Tell If AI Is Hallucinating?
You can tell AI is hallucinating when its answer includes details you can’t trace to a real source. Look for these warning signs:
- Specific details with no source: Exact numbers, names or dates the AI doesn’t attribute
- Citations that don’t open: Links that 404, or papers and cases you can’t find
- Policies you don’t recognize: Prices, terms or rules that aren’t on your website
- Too much confidence: A firm answer to a niche question where a human expert would hedge
- Answers that shift: Ask the same question twice and get two different facts
AI hallucination detection tools
For workflows that run at volume, teams add automated AI hallucination detection:
- Grounding checks: Compare each claim with the documents the AI was given
- A second model as judge: A separate AI reviews the answer for unsupported claims
- Data matching: Check names, prices and totals against your CRM or database
These checks flag likely errors. They don’t replace a person on high-stakes work.
How a Human Check Catches AI Hallucinations
A human check catches hallucinations by putting a person between the AI’s draft and the action. In every example above, one review step would have stopped the damage.
Place the check where a wrong answer costs most:
- Anything sent to a customer, such as replies, quotes and refunds
- Anything with numbers, citations or legal claims
- Anything published, filed or sent to a client
Make reviews fast, or people stop reading. Show the source next to the AI’s claim, and highlight names, amounts and dates. Our human in the loop guide explains how to set up the approval step.
Conclusion
AI hallucinations aren’t rare glitches. They’re a normal side effect of how language models work. Every case above shows the same pattern: A confident wrong answer went out with no one checking it. Ground your AI in real data, watch for the warning signs and put a person where mistakes cost most.
Catch AI Errors Before Your Customers Do With Dscvred Studio
We build AI workflows that answer from your approved data, not from guesswork. Customer replies, figures and anything public go to a named person for approval first. Every check is logged, so you can see what the AI got wrong and fix it at the source.
Explore our AI automation services, or book an intro call to find the riskiest step in your first workflow.
FAQ
Does ChatGPT still hallucinate?
Yes. Newer models hallucinate less, but none are error-free. OpenAI said in 2025 that hallucinations remain a basic challenge for all large language models, including its own.
Which AI hallucinates the most?
It depends on the task and changes with each model update. Public benchmarks such as Vectara’s hallucination leaderboard compare models on the same test. Use them as a guide, then test the model on your own work.
How often do AI hallucinations happen?
It depends on the model and the task. On Vectara’s leaderboard (May 2026 update), models added false details to between 1.8% and 24.2% of document summaries. Rates rise on niche topics and recent events, where the AI has no source to draw on.
