AI hallucinations: why chatbots invent facts
Why language models make things up, which errors are most common, and practical ways to check AI answers before you rely on them.
A hallucination is an AI output that sounds fluent and confident but is false or unsupported. The model is not lying in the human sense. It is doing what it was trained to do: produce a plausible continuation.
Why the model guesses
Language models learn statistical patterns in text. When a question asks for a name, date, citation or quotation, the model predicts the kind of text that usually follows. If the exact fact is missing or weakly represented, a plausible-sounding invention can score better than an admission of uncertainty.
This is especially likely with:
- specific numbers, dates and version names;
- citations, page numbers and legal references;
- local businesses, people and recent events;
- quotes attributed to real people;
- code libraries and API details that changed after training.
A model may also blend two real facts into one wrong statement. The result can be harder to spot than an outright fabrication because parts of it are correct.
Fluency is not evidence
Humans use confidence, detail and polish as cues for expertise. A language model produces all three without an internal fact-check. A long, well-structured answer can be assembled from patterns rather than knowledge.
Treat an AI answer as a draft. The more specific and consequential the claim, the more it needs an independent source.
Checks that actually help
- Ask for sources, then open them. A citation that cannot be opened is not evidence.
- Search the exact claim. If a fact appears nowhere else, that is a warning sign.
- Check primary documents. Use the official site, statute, dataset or paper rather than a summary.
- Ask the model to separate facts from guesses. It may still be wrong, but the answer is easier to audit.
- Use tools for maths and data. A calculator or spreadsheet is more reliable than mental arithmetic in a model.
- Test the same question twice. Inconsistent answers signal uncertainty.
When hallucination is less likely
Retrieval-augmented generation can reduce errors by supplying relevant documents at answer time. Tool use can ground calculations, dates and database lookups. Neither eliminates the problem: the model still decides what to retrieve and how to interpret it.
Constrained tasks with a clear source, such as summarising a document you provide, tend to be safer than open-ended questions about the world. Even then, check that the summary does not add claims that were not in the source.
Designing around the risk
Use AI where a wrong answer is cheap to catch or low-stakes: brainstorming, rewriting, generating test cases, drafting outlines. Add human review before anything reaches a customer, a court filing, a medical decision or a payment.
For high-stakes work, record where each fact came from. If the model cannot point to a source you can verify, treat the claim as unconfirmed. The goal is not to distrust every sentence; it is to match the level of checking to the cost of being wrong.
See what an LLM is for the underlying mechanism and ChatGPT vs AI agents for how tool use changes the risk profile.