Why do AI chatbots make things up?
Why do AI chatbots make things up?
AI chatbots make things up because they learn to produce plausible text, not to check facts, and because the tests used to rank them reward a confident guess over admitting uncertainty. That is the explanation OpenAI researchers gave on Sept. 5, 2025, in Why language models hallucinate. OpenAI defines hallucinations as "plausible but false statements generated by language models."
The longer answer
Where do hallucinations come from?
They start in training. "Language models first learn through pretraining, a process of predicting the next word in huge amounts of text," OpenAI wrote. Spelling and grammar follow patterns a model can learn. Rare facts, such as a particular person's birthday, do not, so the model fills the gap with something that sounds right.
OpenAI gave an example from one of its own researchers, Adam Tauman Kalai: asked for his birthday, a chatbot "gave three different dates, likewise all wrong."
Why don't chatbots just say "I don't know"?
Because most evaluations score only accuracy. OpenAI compares it to a multiple-choice test: a wild guess might be right, while an empty answer earns nothing: "Leaving it blank guarantees a zero." Models graded this way "are encouraged to guess rather than say 'I don't know.'"
The trade-off shows up in OpenAI's own numbers on SimpleQA, a test of short factual questions:
How two OpenAI models handled SimpleQA factual questions
| Result | gpt-5-thinking-mini | OpenAI o4-mini |
|---|---|---|
| Abstained (said it didn't know) | 52% | 1% |
| Correct | 22% | 24% |
| Wrong | 26% | 75% |
The older o4-mini answered slightly more questions correctly, but it was wrong nearly three times as often, because it almost never abstained.
What happens inside the model?
Anthropic found a piece of the mechanism inside Claude. In research published March 27, 2025, it reported that Claude has a circuit that is "on" by default and makes the model say it has insufficient information. When Claude recognizes a "known entity," such as Michael Jordan, a competing feature switches that default off and the model answers.
Hallucinations happen when that switch misfires: Claude recognizes a name but knows nothing else about the person, the "don't know" response is suppressed anyway, and the model produces a plausible but false answer.
How can you reduce hallucinations?
Anthropic's guide for developers lists techniques that also work for everyday users:
- Allow the model to say "I don't know." Explicit permission to admit uncertainty "can drastically reduce false information," Anthropic says.
- Ask for direct quotes from a document before asking for analysis of it.
- Ask for sources for each claim, and have the model withdraw any claim it cannot support.
- Ask the same question more than once. Different answers to the same prompt can be a sign of a hallucination.
OpenAI's proposed fix is aimed at the industry: "Penalize confident errors more than you penalize uncertainty," and give partial credit for honest uncertainty.
What we don't know
No one has a reliable way to eliminate hallucinations. OpenAI argues that accuracy will never reach 100% because some questions cannot be answered, and Anthropic's guide warns that its techniques reduce errors but do not remove them. Rates also vary widely by model and by task, so a figure for one chatbot on one test does not tell you how often another will be wrong.
FAQ
What is an AI hallucination?
A plausible but false statement generated by a language model, in OpenAI's definition, such as an invented date, quote or citation.
Why do chatbots hallucinate?
They are trained to predict likely text rather than verify facts, and most tests reward guessing over saying "I don't know," OpenAI researchers argued in September 2025.
Can AI hallucinations be fixed?
They can be reduced, for example by letting models abstain and asking for sources, but OpenAI and Anthropic both say they cannot be fully eliminated.
How do I stop ChatGPT or Claude from making things up?
Tell it that "I don't know" is acceptable, ask for quotes and sources, and verify any specific fact before you rely on it.
Sources
Cite this
Why do AI chatbots make things up?, New Mind Museum, November 6, 2026. https://newmindmuseum.com/culture/2026/11/why-do-ai-chatbots-make-things-up-hallucinations/
New Mind Museum. (2026, November 6). Why do AI chatbots make things up?. https://newmindmuseum.com/culture/2026/11/why-do-ai-chatbots-make-things-up-hallucinations/
“Why do AI chatbots make things up?.” New Mind Museum, 6 Nov. 2026, https://newmindmuseum.com/culture/2026/11/why-do-ai-chatbots-make-things-up-hallucinations/.
@misc{new-mind-museum-2026-why-do-ai-chatbots-make-things-up,
author = {{New Mind Museum}},
title = {Why do AI chatbots make things up?},
year = {2026},
url = {https://newmindmuseum.com/culture/2026/11/why-do-ai-chatbots-make-things-up-hallucinations/}
}Free to quote and cite with a link to this page.