In plain English
An AI hallucination is an output that appears plausible but is unsupported, inaccurate, or fabricated. Language models are optimized to produce likely text, not to guarantee that every statement corresponds to evidence.
The main ideas
Plausibility is the objective
A model can compose a convincing answer from familiar language patterns even when the underlying detail is absent or wrong.
Missing context invites invention
When a prompt assumes a false fact or omits necessary evidence, the model may complete the pattern instead of stopping.
Errors can be mixed with truth
A response may contain accurate general information beside an invented date, quotation, case, or source.
Tools help but do not eliminate risk
Retrieval, search, calculators, and citations can improve grounding, but the final explanation still requires appropriate checking.
How it works
- 1
Identify checkable claims
Mark names, dates, numbers, quotations, legal rules, medical statements, and other details that could be verified.
- 2
Trace each claim to evidence
Open the cited or supplied source and confirm that it actually supports the statement.
- 3
Use an independent source
For important claims, compare with an authoritative source not generated from the same response.
- 4
Correct the record
Remove unsupported details, document uncertainty, and do not preserve a fabricated claim merely because it sounds useful.
Where you may see it
Draft review
Editors verify generated copy against source notes before publication.
Research assistance
AI can suggest questions and search terms, while the researcher reads and cites original material.
Document summarization
The reviewer compares the summary with the document, especially exceptions and numerical details.
Decision support
People inspect assumptions, inputs, and evidence before treating an AI output as a basis for action.
Important limits
- Asking the same model to confirm itself is not independent verification.
- A fabricated citation can look professionally formatted.
- Confidence language is not a calibrated probability.
- Current information may be missing unless the system has a reliable live source.
A practical reader checklist
- Separate claims from commentary.
- Open every important citation.
- Recalculate numbers with a trusted tool.
- Use primary or authoritative sources when consequences matter.
Key takeaway
An AI hallucination is an output that appears plausible but is unsupported, inaccurate, or fabricated. Language models are optimized to produce likely text, not to guarantee that every statement corresponds to evidence.
Verification should match the claim
Not every sentence requires the same level of investigation. A stylistic suggestion may need only editorial judgment, while a quotation, medical statement, financial figure, legal rule, or safety instruction requires authoritative evidence. Start by identifying claims that can change a decision or create harm. Confirm them in the original document, official database, qualified source, or trusted calculation. Check that the cited source says what the generated text claims it says. When evidence is unavailable, the honest correction is to remove the detail or label the uncertainty—not to preserve it because it sounds plausible.