Start here
AI Hallucination in plain language
An assistant may invent a citation, add an unmentioned deadline or combine facts from different documents incorrectly. These failures matter because a polished answer can hide the difference between evidence and a likely-sounding completion.
Evaluation depends on the task. A fictional story allows invented details; an invoice extraction task does not. Define which claims must match a source and what the system should do when evidence is absent or conflicting.
How it works
Locate the claim
Break an answer into checkable assertions, such as names, dates, quantities and commitments.
Find support
Inspect the cited source and relevant passage. A working link alone is not evidence for the claim.
Correct the workflow
Record whether the failure came from missing context, retrieval, interpretation or generation. Adjust and retest the relevant component.
Illustrative example · Fictional data
An invented delivery date
- Input
- Fictional source: “The replacement has been approved. Dispatch will be confirmed separately.” Question: when will the replacement arrive?
- Process
- Ask for an answer using only the source and inspect any date or delivery promise in the response.
- Expected output
- Expected answer: the source does not state an arrival date; dispatch is still to be confirmed.
- How to check it
- “It will arrive in three days” is unsupported. Asking the model whether it is confident does not verify the date; the source must supply it.
Where it helps
- Reviewing customer-facing drafts for invented promises.
- Checking citations and summaries against their actual sources.
- Designing evaluation cases where the correct response is to acknowledge missing information.
Hallucination, bias and missing knowledge
- Hallucination
- An unsupported or false generated assertion.
- Bias
- A systematic skew in behaviour or outcomes. It can occur even when individual facts are correct.
- Missing knowledge
- An information gap. A system can handle it appropriately by asking for context or acknowledging uncertainty.
Why fluent generation can be unsupported
A language model predicts plausible continuations; that objective does not require each sentence to be linked to a verified source. If evidence is missing, ambiguous or outside the available context, it can still produce a plausible name, date, quotation or explanation. A request to be confident can change presentation without improving evidence.
Retrieval supplies context but can fail through missing documents, weak ranking or incorrect interpretation. A citation may point to a real document that does not support the sentence. Test support claim by claim and allow the system to abstain or ask for information instead of forcing a complete answer.
Review changes with the stakes
For a brainstorming draft, a reviewer may simply reject weak ideas. For medical, legal, financial or safety-sensitive content, verify the important claims against authoritative, applicable sources and involve a qualified reviewer. Do not turn a model’s self-reported confidence into a safety threshold.
Keep the source, date, exact claim and review outcome in the evaluation record. An unsupported statement should be corrected or removed from the draft, while legitimate uncertainty should remain visible. Hallucination and bias can coexist, but factual support tests do not replace subgroup fairness checks.
Related reading and learning pathways
Common pitfalls
- Lower randomness settings do not guarantee truthful answers.
- Retrieval can supply useful evidence but also outdated or irrelevant passages.
- Model self-checks can help a review process but are not independent verification.
Put the idea to work
Try it yourself
Use the replacement message above. Ask for a short customer reply, then highlight every factual claim. Mark each as supported, contradicted or absent from the source. Revise the reply until every promise has support.
Common questions
Can hallucinations be eliminated with a better prompt?
A clearer prompt can help, but it does not guarantee correctness. Use task-specific checks and evidence, particularly for consequential answers. Grounding answers in retrieved sources, asking the model to say when it is unsure and adding human review for high-stakes outputs reduce the risk further.
Are citations proof that an answer is right?
No. Check that the source exists, is suitable and supports the precise claim. A real citation can still be irrelevant. Models can invent references or cite real ones that say something different. Open the source and confirm the specific statement before relying on it.
Is every wrong AI output a hallucination?
Terminology varies. In practice, distinguish unsupported generation from calculation, retrieval, data and software errors so you can fix the actual cause. A wrong total may come from bad input data, a failed retrieval may come from search settings, and a broken integration may be a software bug. Each needs a different fix.
Your next step
Continue learning
Connected concepts
Sources and further reading
Examples on this page are teaching exercises using fictional information. They are not client case studies or measured outcomes.