An AI hallucination is an answer that sounds completely true and isn’t. You get them because a large language model (LLM) predicts an answer rather than retrieving one. When the training text behind your question is thin, the model fills the gap with something that reads correctly, and it comes back in the same confident tone as an answer that happens to be right.
What is an AI hallucination?
An AI hallucination is an answer that sounds completely true and isn’t.
Ask a model what Einstein said about creativity and it hands back a quote. “Creativity is intelligence having fun.” It sounds like him. Fact check it and he never said it. Nothing in the wording of the response separates that quote from a real one.
AI hallucinations cluster in four predictable places.
- Statistics and numbers. Models exaggerate or invent figures mid-sentence.
- Citations and sources. A document, a clause or a page reference that does not exist.
- Quotes and attributions. The Einstein line.
- Names and addresses. A supplier that does not exist, a company name that is close but wrong, a service address nobody has ever driven to.
Any kind of data the model thinks it knows, it might not know.
What causes AI hallucinations?
The first cause is what we call the confidence gap. Plot the model’s real accuracy against how common the topic is, and accuracy sits high on common topics and drops away as the topic gets rare and niche. Confidence stays flat across the whole range. To cover the gap, the model creates new data.
| Your question | What the training text holds | What the model does |
|---|---|---|
| How does a heat pump move heat? | Textbooks, manuals, forum threads | Recalls a well-worn answer |
| What is the part number on this discontinued rooftop unit? | Almost nothing | Invents a plausible number |
| What does this county require at inspection? | Almost nothing | Invents a plausible rule |
All three answers arrive in the same tone at the same speed, and one of them sends a tech out with the wrong part.
The second cause is the age of the training text. GPT-3.5, the model that put ChatGPT in front of everyone in 2023, was trained on data up to September 2022 and knew nothing after that date. Models today notice the edge of what they know and can call a tool to go and search, but every model still has a cutoff, and the price list you changed last week sits outside it until you hand it over.
Why a database gives you the right answer and a model does not
Query your CRM for a tenant’s address and you get the row that is there. Data in a database does not change unless someone changes it. Query a model and you are dealing with a black box. It does not know what you will ask, and it does not know what the output will be until you ask the question. The same question asked twice can come back two different ways.
Can you eliminate AI hallucinations?
No. Prediction is the same property that makes a model useful on questions nobody has ever written a database query for, so patching out the hallucination means losing the thing you came for. What you can do is shrink the number of places a wrong answer can survive, and check the output before anyone acts on it.
Do all AI models hallucinate?
Yes. Anything that predicts its output hallucinates, whichever provider’s logo is on it. Newer models are better at recognising the edge of their knowledge and reaching for a search tool, which is why the same question can come back sourced from one model and invented by another. Plan around the guess, because the answer looks identical either way.
How to reduce AI hallucinations
Four controls, roughly in order of how much they buy you.
- Paste in the source material. Transcripts, documents, the actual invoice or quote. A ten-page contract is nearly 7,000 tokens, which any current model reads in one go. Ask a model what your cancellation terms are and you get an answer shaped like a cancellation clause. Paste in the contract and ask again, and you get the notice period your customer signed.
- Connect it to live search. Third-party tools give the model real-time web search and real-time documentation search, which handles the training cutoff. Context7 does live documentation search. xAI and NIA MCP do live web search. Most of them plug in as MCP servers.
- Make it cite sources. Ask for a citation on every claim, then open one or two of them. Citations that point at nothing are hallucinations too, and they are fast to catch.
- Break the task into smaller asks. Give the model one stage at a time and verify each stage before it moves on. A long unsupervised run takes a wrong figure from step two and carries it through steps three, four and five.
With all four in place you will still get hallucinations. Fewer of them, and in places you are looking.
Pick your controls by what a wrong answer costs. A draft reply with a wrong number costs you an apology. A price on a customer quote, a date on a permit and an address on a dispatch turn into actions, and an action taken on invented data is expensive to reverse. Anywhere the output becomes something a crew or a customer acts on, put the source material in front of the model and make it show where the answer came from. The Einstein quote sounded right too. Later in this track we get into how AI agents are built to hold up under this, with the live-search tools running against real work.
