An AI hallucination is an output that sounds plausible but contains false or unsupported information. It might be an invented reference, an incorrect feature description, or a detail that was never in the material you provided.
You don’t need to recognize every possible error in advance. A useful habit is to break an answer into claims and ask what supports each one. A confident tone and a working link are starting points for review, rather than evidence by themselves.
For the underlying concepts, see our introduction to AI.
A small example you can verify yourself
Here is an invented announcement for a community workshop:
The workshop has space for eight participants. Register by replying before noon on Thursday. The session takes place on Saturday. Bring your own fabric; shared scissors will be available. The venue is still to be confirmed.
Now imagine a summary that says:
Join Saturday’s free workshop at the community center. Ten places are available, and all materials are provided. Register on Thursday.
This is a deliberately constructed error example, not a record of a particular model’s performance.
| Claim in the summary | What the announcement actually says |
|---|---|
| Saturday session | Supported |
| Free workshop | No price information is given |
| Community center | Venue is unconfirmed |
| Ten places | Eight places |
| All materials provided | Bring fabric; scissors are shared |
| Register on Thursday | Reply before noon on Thursday |
The mistakes are ordinary enough to slip through a quick read. Some facts changed, some details appeared from nowhere, and a deadline lost a condition that matters.
Why does AI make things up?
A language model generates an answer using patterns and the context available to it. That process does not automatically verify every claim. Missing information, unclear requests, and unreliable source material can all contribute to problems.
Tools that retrieve web pages or documents can provide evidence, but their use doesn’t establish that every resulting sentence is supported. AWS’s overview of hallucinations describes this broader reliability problem.
In practice, “Where did this detail come from?” is more useful than trying to decide whether the system intended to deceive you.
A four-step fact-checking workflow
1. Separate the claims
Split a paragraph into statements you could check. Pay attention to names, quantities, dates, product features, and instructions.
For example, “All new users can use this feature for free” contains several claims: who qualifies, which feature is included, and what it costs. A source may support one part while leaving another unanswered.
2. Distinguish facts from inferences and advice
A fact needs evidence. An inference should identify its assumptions. Advice should explain the goal it serves.
“Confirm that the venue has power outlets” is advice. “The venue has power outlets” is a factual claim. An answer should not silently turn the former into the latter.
3. Inspect the original evidence
Open the cited page and locate the supporting passage. Check who published it, when it applies, and whether it covers the relevant version or region.
For your own documents, compare the answer with the document itself. Recalculate important numbers. A real publication can still be misquoted, and a real page can still be irrelevant to the sentence citing it.
4. Keep unresolved points visible
“The source doesn’t state the fee” is a useful result. Filling that space with a plausible price would hide a task you still need to complete.
Match the review to the consequences of using the answer. A draft can help you organize information, but the facts you act on still need confirmation.
A prompt that makes checking easier
Answer using only the material below. Separate the result into “Explicitly stated,” “Inferences,” and “Not established.” For each factual claim, identify a supporting section or short passage. Don’t invent prices, dates, quantities, or features. If the material doesn’t support a claim, say so. Source material: [paste text].
Anthropic’s guidance on reducing hallucinations recommends allowing uncertainty and grounding answers in source passages. These techniques make review easier; they don’t eliminate the need to check the passages and references yourself.
What doesn’t count as independent verification?
Asking “Are you sure?” may produce a different answer without supplying evidence. Getting the same answer from several AI tools can reveal agreement, but the tools may share the same mistaken source.
A self-reported confidence percentage is also insufficient on its own. Unless it comes from a properly evaluated and calibrated system, don’t treat it as the probability that a particular statement is true.
For recurring tasks, define your own review points. Invitations need correct logistics. Summaries need conditions and exceptions preserved. Research notes need sources that support the actual claims.
Read how to summarize documents with evidence checks for a worked workflow, or how to write clearer AI prompts if your requests leave too much information unspecified.