GuidesAI reliability
AI hallucinations: why AI makes things up, and how to stop it
Why a model makes things up
A language model completes a text with the most likely continuation. When information is missing, the most likely continuation is an answer that looks right. It is not lying, it is filling the gap.
What causes hallucinations
What prevents them
The method, in five points
- 1Provide the sourceThe useful passages, retrieved from your documents (RAG).
- 2Allow ignorance“If the answer is not in the documents, say so.”
- 3Require citationsEvery claim points to the passage that supports it.
- 4Check in codePrices, dates, references: compared with the database before display.
- 5TestA set of trick questions, replayed on every change.
Where to be most careful
| Use | Risk | Safeguard |
|---|---|---|
| Customer service | A made-up promise | Answers limited to approved content |
| Figures, prices | A wrong amount | Value read from the database, never generated |
| Legal, health | A binding error | Systematic human review |
See also AI chatbot for a website and AI agents on real data.
Frequently asked questions
- Can hallucinations be removed completely?
- No. You make them rare and detectable: answers backed by a source, code checks on figures and references, human review on anything binding.
- Is lowering the temperature enough?
- No. It makes answers more consistent, not more true. A model can repeat the same mistake with confidence.
- Do newer models hallucinate less?
- Generally yes, but the first lever stays the same: give it the information rather than rely on its memory.