AI Hallucinations Explained: Why Chatbots Make Things Up (and How to Catch Them)

Ask an AI chatbot for a list of sources and you might get perfectly formatted citations — to books that do not exist. Ask about a small company and you might get a confident history that is entirely invented. This behavior is called hallucination, and anyone who uses AI tools needs to understand it.

What is an AI hallucination?

A hallucination is when an AI system produces information that sounds plausible but is false, unsupported, or made up — and presents it as fact. It is not lying in the human sense; the model has no intention to deceive. It is a side effect of how these systems are built.

Why does it happen?

1. Models predict likely text, not true text

A large language model generates answers by predicting the most likely next word based on patterns from its training. Most of the time, likely and true line up. But when the model lacks solid knowledge, it still produces the most likely-sounding continuation — which can be fiction dressed as fact.

2. No built-in fact database

Unless connected to search or documents, the model is not looking anything up. It is reconstructing information from compressed patterns, a bit like recalling a conversation from years ago: the gist may be right while the specific names and numbers drift.

3. Pressure to answer

Assistants are trained to be helpful, and historically “I don’t know” was underrepresented in examples of good answers. Newer models are better at admitting uncertainty, but the tendency to fill gaps remains.

4. Rare and recent topics

The less often something appears in training data — niche subjects, small businesses, local news, recent events — the weaker the model’s knowledge and the higher the risk.

Where hallucinations are most common

• Citations and quotes — fake paper titles, wrong authors, invented quotations.
• Specific numbers — statistics, dates, prices, and percentages.
• Little-known people, places, and companies.
• Legal and medical specifics — case names, dosages, regulations.
• Recent events after the model’s training cutoff.
• Technical details — functions or settings that “should” exist in a piece of software but don’t.

A practical checklist for catching hallucinations

1. Be suspicious of precision. Exact figures and specific names are where errors hide. Verify anything you will publish, sell, or act on.
2. Check every citation. Search for the title and author. If you cannot find it, assume it does not exist.
3. Ask for sources — then open them. Tools with web search can give real links, but still click through and confirm the page says what the AI claims.
4. Ask “How confident are you?” Models will often flag shaky claims when prompted directly.
5. Re-ask differently. Ask the same factual question in a new chat. If answers disagree, treat both as unverified.
6. Give it the facts. Paste the document, data, or article you want it to work from and tell it to answer only from that material. This is one of the most effective ways to reduce made-up content.
7. Use a primary source for anything important. Official websites, original studies, and government pages beat any summary.

Prompts that reduce hallucinations

• “If you’re not sure, say so rather than guessing.”
• “Answer only using the text I’ve provided. If the answer isn’t there, tell me.”
• “Separate what you know confidently from what you’re inferring.”

For more techniques, see our prompt writing basics.

Should you stop trusting AI?

No — but trust it the right way. AI is excellent for drafting, brainstorming, explaining concepts, restructuring text, and working with material you provide. It is risky as an unchecked source of facts. Think of it as a brilliant, fast assistant who occasionally misremembers: enormously useful, as long as someone checks the important details.