Learn · Verification
AI-generated content needs checking.
A convincing sentence isn't evidence that it's correct. AI-generated text can misdescribe what is present, invent missing details, confuse observation with inference, and produce references that do not support its claims.
The purpose of verification isn't to make an output sound more polished. It's to establish which parts are supported, which need correction and which should not be published.
Start with the original evidence
Keep the original image, video, document, interview or trusted source available while reviewing. Don't use the generated answer as the only source against which to check itself.
- Separate claims. Break the output into statements you can verify rather than evaluating the paragraph as a whole.
- Check what can be observed. For an image or video, inspect what is visible or audible. For a factual claim, trace it to a reliable source.
- Mark uncertainty. Distinguish what is confirmed, what is inferred and what cannot be established.
- Check omissions. Look for important information missing from the output, not only incorrect additions.
- Check the effect on the reader. Ask whether an error could change someone's understanding, decisions or ability to complete a task.
- Revise and verify again. Re-check the final version, not only the first AI draft.
Warning signs worth looking for
- Invented detail: objects, actions, speech, intentions, causes or timelines that the source doesn't establish.
- Overconfidence: wording such as “clearly,” “definitely” or “because” when evidence is limited.
- Wrong specificity: precise numbers, identities, colours or locations that haven't been verified.
- Confused sequence: events described out of order or merged together.
- Unsupported citations: a real-looking reference that is missing, irrelevant or does not support the claim.
- Missing context: a technically accurate statement that is misleading without an important qualification.
- Unnecessary interpretation: assumptions about emotions, ability, diagnosis or intent from limited observations.
- Accessibility regressions: unclear reading order, poor descriptions, ambiguous instructions or jargon introduced during rewriting.
A simple verification record
- Claim
- What does the generated text say?
- Source
- What original evidence can establish it?
- Assessment
- Confirmed, corrected, uncertain or unsupported?
- Action
- Keep, rewrite, qualify, remove or request human review.
- Remaining limitation
- What still cannot be verified?
How much checking is enough?
The risk determines the level of review. Low-stakes drafting may need a careful read against its source. Accessibility instructions, descriptions, medical or legal information, safety advice and consequential decisions demand stronger independent verification and appropriate expertise. If a material claim cannot be checked, don't present it as established fact.
AI can assist the review process, but asking another AI to agree with the first output isn't independent verification. Human review also isn't automatically accurate: give reviewers the source, enough time, and permission to challenge the output.
Try this with accessibility descriptions
Image and video descriptions are a particularly clear example because errors can replace information someone relies on. Review how to check a generated accessibility description.
What a review does not prove
A checked output isn't guaranteed to be complete or appropriate for every person or context. Record who reviewed it, against what source, and what remained uncertain when that information matters.