Knowledge · Ira Berg, 12 November 2026
Why does ChatGPT make things up? AI hallucinations explained
In short: AI models predict the word that fits best. Whether it is true is not checked along the way. That is why wrong answers often sound just as confident as right ones. Experts call this hallucinating.
What a hallucination is
When an AI claims something that is not true but sounds certain, it is called a hallucination. It can be wrong dates, made up quotes, books that never existed or links that lead nowhere. The tricky part: the answer is usually fluent and convincing.
Why it happens
Language models like ChatGPT, Gemini and Claude have learned from huge amounts of text which words often follow each other. For every answer they predict, piece by piece, what is most likely to come next. That works surprisingly well, but it is something different from looking things up. A model has no built in sense of whether it really knows something. If knowledge is missing, it fills the gap with something that sounds plausible.
It happens especially often with:
- Sources and quotes: titles, authors and page numbers are often invented.
- Rare topics: the less there is about something on the internet, the more the model guesses.
- Recent events: what happened after its training, the model only knows if it searches the web.
- Numbers and calculations: longer calculations sometimes go wrong without the AI noticing.
How to notice it
- Ask for the source and open it yourself. If the link does not exist or says something else, be careful.
- Ask the same question again or phrase it differently. If the answer changes a lot, the AI is not sure.
- Check important facts in a second source, such as your textbook or a reputable website.
- Be suspicious when an answer sounds very precise, for example exact percentages without an origin.
A detailed guide is in Checking AI answers in 3 steps.
Why this matters for your learning
If you simply take over AI answers, you also take over their mistakes. If you read, question and check them, you even learn more, because you really engage with the topic. That is exactly what Monkey Score measures: whether you just copy answers or work with them yourself. Thinking along earns you golden bananas.
FAQ
Does the AI lie on purpose?
No. An AI has no intention. It produces the most likely answer, and sometimes the most likely answer is simply wrong.
Are hallucinations getting less common?
The models are getting better, and many now search the web and cite sources. But mistakes will not disappear completely, so checking remains important.
Prompt to try
Copy this straight into ChatGPT, Gemini or Claude:
“Answer my question and then tell me honestly which parts of your answer you are unsure about and what I should check myself: [paste question]”