You ask a chatbot for a source and it gives you a neat citation: author, year, journal, page numbers. You search for it and the paper does not exist. The model was not lying in any human sense. It was doing exactly what it was built to do, which is the part most people find surprising.
This guide explains why AI "hallucinates", the questions where it is most likely to happen, and a short routine for catching it.
What an AI Hallucination Actually Is
A hallucination is an answer that is **fluent, confident and wrong**. Typical examples:
- A book, paper or court case that was never written
- A quote attributed to a real person who never said it
- A statistic with a precise number and no real origin
- A software function or setting that does not exist
- A summary that adds a detail the original text never contained
The confidence is the dangerous part. A wrong answer delivered with hesitation gets checked. A wrong answer formatted like an encyclopedia entry gets pasted into an email.
Why It Happens
Large language models generate text one piece at a time by predicting what is most likely to come next, based on patterns learned from huge amounts of writing. That is why they are so good at tone, structure and explanation.
The catch is that **"likely to come next" is not the same as "true"**. A few things push the two apart:
- No fact-checking step. The model does not look up each sentence in a database before writing it. It produces what fits the pattern
- Gaps in training data. For rare topics, the model has seen little or nothing, but it can still produce text in the right shape
- Training rewards answering. Models have historically been rewarded for helpful-looking answers more than for saying "I do not know", which nudges them toward guessing
- Patterns are easy to imitate. A citation has a very regular format. Producing something that looks like one is easy; producing a real one requires actually knowing it
- Knowledge cutoff. Without a web search tool, a model knows nothing after its training data ends, but questions about recent events can still get an answer
Newer models hallucinate less, and many now search the web or admit uncertainty more often. None of them are immune.
Where Hallucinations Are Most Likely
| Question type | Risk | Why |
|---|---|---|
| Rewrite, summarise or translate text you pasted | Low | The facts are in front of the model |
| Explain a well-known concept | Low to medium | Widely covered in training data |
| Exact numbers, dates, prices | High | Small details are easy to get slightly wrong |
| Quotes and citations | Very high | Format is easy to fake, content is hard to recall |
| Niche people, small companies, local rules | High | Little training data to draw on |
| Recent events (without web search) | Very high | Outside the training data entirely |
| Legal, medical, tax specifics | High stakes | Errors here cost real money or health |
The rule of thumb: **the more specific and checkable the claim, the more you should check it.**
A Five-Step Checklist for Catching Made-Up Answers
1. Separate the useful from the factual. Structure, wording, brainstorming and explanations are where AI shines. Facts that will be published, sent or acted on are the parts that need checking.
2. Open every source. If the answer cites a paper, article or link, open it. Confirm it exists and that it actually says what the AI claimed. A real source quoted wrongly is almost as common as a fake one.
3. Ask again in a fresh chat. Ask the same factual question in a new conversation. A real fact tends to stay stable. A hallucinated one often changes: a different year, a different author, a different number.
4. Ask the model to show its uncertainty. Prompts like "Which parts of this answer are you least sure about?" or "Say if you do not know" often surface the weak spots. Treat this as a hint, not proof.
5. Verify high-stakes facts at the primary source. Prices on the vendor's page, laws on the government site, dosages from a pharmacist or clinician. Not a second chatbot.
Prompts That Reduce Hallucinations
You cannot prompt your way to zero errors, but you can lower the rate:
- Give it the material. "Using only the text below, answer..." keeps the model working from facts you supplied
- Allow "I do not know". "If the text does not say, reply that it is not stated" removes the pressure to invent
- Ask for quotes from your source. Requiring an exact supporting quote from the pasted text makes inventions easier to spot
- Break big questions down. A short, focused question is easier to answer correctly than a sprawling one
If you use AI for daily work, our list of real AI productivity workflows leans on exactly these tasks where the model works from your material rather than its memory.
Is One Chatbot More Reliable Than Another?
Models differ, and the gap changes with every release. Web search, newer training data and better calibration all help. Our ChatGPT vs Claude vs Gemini comparison covers the current strengths of each. But the checking habit matters more than the brand: every current model can produce a convincing wrong answer.
A related point: tools that claim to detect AI writing have their own accuracy problems, covered in do AI detectors work.
Using Generai With This in Mind
Generai is an AI chat and image app for iPhone. It is well suited to the low-risk, high-value side of the table above: drafting, rewriting, summarising text you paste in, brainstorming and explaining concepts.
What to keep in mind:
- It can hallucinate too. Every language model can. Use the checklist above for any fact you plan to rely on
- It is not a professional adviser. For legal, medical or financial decisions, use the AI to prepare your questions, then get the answers from a qualified person
Used that way, a chatbot saves real time without becoming the source of your next mistake.