AI is very good at sounding sure of itself.
Like George Clooney in Ocean’s Eleven, it smoothly explains the situation and gives you the ‘facts’.
Depending on the context, the ‘facts’ that AI shares with you might be a quote, a statistic, or an explanation delivered in confident tone.
Whether the ‘facts’ are true or completely made up is for the human to figure out.
This behaviour is called an AI hallucination: a generated AI answer that is confident-sounding and false.
So why does AI make things up? And can you trust anything it tells you?
In this article, I’ll explain why AI hallucinates in plain English, share some of the best-known AI hallucination examples, and walk through what the latest research from OpenAI and Anthropic says about fixing AI hallucinations.
We’ll also talk through how to reduce the risks of AI hallucinations in your daily life.
Why Does AI Make Things Up?
Short answer: AI makes things up because it predicts the most likely next word instead of looking up facts. And during training, it was rewarded for guessing rather than saying “I don’t know.”
So when AI doesn’t know an answer, a guess that sounds right is often what comes out. That guess is called an AI hallucination.
What Is an AI Hallucination?
An AI hallucination is a confident answer that’s false. It might be a fact, a quote, a statistic or a source that doesn’t exist.
The tricky part is that hallucinations look exactly like the true answers. They come out sure, smooth and well written, with nothing to warn you.
Some researchers prefer the word “confabulation,” which means filling a gap in memory with something that sounds plausible. That’s a pretty good description of what’s going on.
What Are Some Real AI Hallucination Examples?
AI hallucinations have moved well beyond funny chatbot mistakes. Here are four that made the news.
- Fake court cases. Lawyers keep filing briefs with case law that AI invented. Researcher Damien Charlotin’s AI Hallucination Cases database had logged 2,184 court and tribunal decisions involving hallucinated content, such as fake citations, as of October 2026.
- The airline chatbot. Air Canada’s website chatbot told a grieving passenger he could claim a bereavement discount after booking. That policy didn’t exist. Air Canada argued the chatbot was responsible for its own actions. A Canadian tribunal called that “remarkable” and ordered the airline to pay.
- The summer reading list. In May 2025, the Chicago Sun-Times ran an AI-made summer reading list. Ten of the 15 books didn’t exist, including a novel credited to Isabel Allende.
- Asking about yourself. OpenAI researcher Adam Kalai asked a popular chatbot for the title of his own PhD dissertation. It gave three different answers. All of them were wrong.
A made-up book title is funny. A made-up court case in a legal filing is a much bigger deal.
How Does AI Decide What to Say?
AI language models work like autocomplete with a lot of confidence. They predict the next likely word, one AI token at a time, based on patterns in their training data.
That works brilliantly for anything with a pattern. Spelling, grammar and well-known facts show up millions of times in the training data, so they’re easy to predict.
One-off facts are different. Someone’s birthday, the title of a thesis or the exact wording of a court ruling might appear once in the training data, or not at all.
There’s no pattern to learn, so the model fills the gap with something that sounds right. OpenAI’s researchers point out that spelling mistakes fade as models get bigger, while random facts like birthdays keep tripping them up.
Why Does AI Hallucinate Instead of Saying “I Don’t Know”?
AI hallucinates instead of admitting it doesn’t know because it was trained and tested like a student sitting a multiple-choice exam. A guess might score a point. A blank answer always scores zero.
That’s the main argument in OpenAI’s September 2025 paper, Why Language Models Hallucinate. As the authors put it, “Leaving it blank guarantees a zero.”
Most AI tests only count how many answers are right. They don’t take points off for confident wrong answers.
If a model guesses someone’s birthday, it has a 1-in-365 chance of being right. If it says “I don’t know,” it’s guaranteed to score nothing. Over thousands of questions, the model that guesses looks better on the scoreboard.
In other words, we accidentally trained AI to bluff.
What Happens When AI Is Allowed to Say “I Don’t Know”?
When AI is allowed to hold back, its wrong answers drop sharply and its accuracy barely changes. OpenAI’s own test results show the difference.
| SimpleQA test | o4-mini (the guesser) | GPT-5 thinking mini (the careful one) |
|---|---|---|
| Said “I don’t know” | 1% | 52% |
| Right answers | 24% | 22% |
| Wrong answers | 75% | 26% |
The careful model got two fewer questions right, and about a third as many wrong.
Smarter doesn’t automatically mean more truthful, either. In April 2025, TechCrunch reported that OpenAI’s o3 reasoning model made things up on 33% of questions about people, roughly double the rate of the older o1. Its sibling, o4-mini, hit 48%.
What’s Going On Inside the AI When It Hallucinates?
Anthropic’s researchers found that Claude has a built-in “I can’t answer that” response that’s switched on by default. When Claude recognizes a name or topic, a “known entity” signal switches the default off, and it answers.
In their March 2025 study, Tracing the Thoughts of a Large Language Model, they asked Claude about basketball star Michael Jordan. The switch turned off and Claude answered. Then they asked about “Michael Batkin,” a made-up name. The switch stayed on, and Claude declined.
Hallucinations happen when that system misfires. If a name feels familiar but Claude doesn’t know anything else about the person, the “I know this” signal can still flip the switch. Then the model fills the gap with a plausible answer.
It’s a lot like someone at a party nodding along to a name they half recognize.
The researchers could even trigger it on purpose. By switching on the “known answer” signal, they got Claude to claim Michael Batkin plays chess.
Will AI Ever Stop Hallucinating?
AI will probably never reach zero hallucinations, but it can get much better at knowing when to hold back. Researchers disagree on how far that goes.
- Built into the math. A paper by Xu, Jain and Kankanhalli argues hallucination is inevitable, because no language model can learn everything. It’s a theoretical argument, shared as a preprint on arXiv.
- Fix the scoreboard. OpenAI argues hallucinations aren’t inevitable, because models can choose not to answer. Their fix is to change how AI is graded, so a confident error costs more than “I don’t know.”
- Some questions have no answer. Even OpenAI says accuracy will never hit 100%, because some facts can’t be known from the data.
- Better, not perfect. On Vectara’s hallucination leaderboard (May 2026), the best models still add something false to about 3% of document summaries, with the source text right in front of them. Some popular reasoning models slip on more than 20%.
For now, the safest assumption is that any AI tool can make something up, including the newest ones.
How Do You Catch AI Hallucinations?
The best way to catch AI hallucinations is to check anything that matters, and set the AI up so guessing isn’t its only option. Three habits go a long way.
- Ask for sources, then click them. If a link doesn’t open, or the page doesn’t say what the AI claims, don’t trust it. This matters most for statistics, quotes, legal or medical info, and anything with a person’s name attached. (As we found in Is AI making us dumber?, the more people trust AI, the less they tend to check it.)
- Let it say “I don’t know.” Add a line like “If you’re not sure, just say so. I’d rather have no answer than a wrong one.” It gives ChatGPT or Claude permission to hold back.
- Give it the facts first. Paste in the document, report or web page, and ask it to answer only from that. AI is far more reliable working from material in front of it than from memory.
That last habit is how the research for the show works too. Every stat on our slides links back to a source we opened and read.
Treat AI Like a Brilliant, Overconfident Friend
The more you learn about why AI makes things up, the less mysterious it gets. It’s a word-predicting machine that was rewarded for sounding sure.
That’s a fixable problem, and the companies building these tools are working on it. Models that say “I don’t know” more often are a very good sign.
Until then, treat AI like a brilliant friend who hates admitting they’re stumped. Great company, great ideas, and always double-check before you quote them in court.
– Daniel Midson-Short
Leave a Reply