Recognising hallucinations and reviewing AI-generated work

Why hallucinations are hard to spot

A hallucination looks exactly like a correct answer: same fluency, same confidence, often with convincing detail. People also tend to trust text that is well written and agrees with what they expect (automation bias and confirmation bias). That combination is why errors slip through.

Warning signs

Be especially careful when output contains:

• Precise numbers, dates or percentages without a source you provided.
• Citations, quotes, case names or URLs — check each one exists and says what is claimed.
• Named people, products or organisations doing or saying specific things.
• Niche or recent topics, where training data is thin or out of date.
• Overly neat answers to genuinely uncertain questions.
• Answers that change when you rephrase or re-run the same prompt.
• Explanations that sound logical but skip a step you cannot verify.

None of these proves an error; each means "verify before use".

A proportionate review process

Match the review to the stakes:

A review checklist

Before using AI-assisted work:

1. Accuracy — every factual claim checked or removed.
2. Sources — citations exist and support the claims.
3. Completeness — nothing important missing; caveats preserved.
4. Fairness — no stereotypes, one-sided framing or excluded groups.
5. Fit — right audience, tone and format.
6. Rules — complies with your organisation's, school's or platform's policies, including any disclosure requirement.
7. Ownership — you can explain and defend every part of it.

Bias

Models learn patterns from human-written text, including stereotypes and imbalances. Bias can appear in examples chosen, assumptions about people's roles, or which perspectives are presented as normal. Ask for diverse examples, check who is represented, and review outputs that describe or affect people with extra care.

Prompt injection, revisited

When a model processes text from elsewhere — web pages, emails, documents — that text may contain instructions designed to manipulate it ("ignore your instructions and…"). Delimiters and clear instructions reduce the risk; they do not remove it. Do not let AI workflows that read untrusted content take consequential actions (sending, deleting, paying) without human review.