AI misinformation risk checklist
AI writes fluently, which is exactly what makes its mistakes dangerous — a confident, well-structured paragraph can contain an invented statistic, a fake citation, or a fact that was true two years ago and isn’t now. This checklist forces the verification steps people skip when content “looks finished,” and gives you a clear risk level before you hit publish.
The five verification areas
Factual claims — every specific, verifiable statement in the content needs to be traced back to a source you can check. Statistics, percentages, named events, legal findings, scientific claims, and historical facts are all in scope. “Studies show” followed by a number requires knowing which study and confirming the number. The model may have generated a plausible number from surrounding context rather than a real data point.
Currency — AI models have training cutoffs, and facts go stale. Laws change, prices change, software versions change, people change roles. Any claim about current state — “as of today,” “the current rate,” “the latest guidance” — is particularly vulnerable. For time-sensitive content, check whether the model’s training data predates any relevant changes.
Citations — language models sometimes invent citations that look plausible: a study by named researchers at a recognisable institution in a credible journal with a year and volume number, none of which is real. The check is simple: open the citation and confirm it exists and says what the text claims. A citation that cannot be found is a hallucinated citation; it must be removed, not left with a “may be inaccurate” caveat.
Framing — misinformation is not always false facts. Selectively true information, cherry-picked statistics, missing context, and implied causation where there is only correlation can all mislead without containing a single false statement. This check asks whether the content gives a reasonably complete picture of the issue or whether it presents a partial view as if it were the whole.
Accountability — is it clear who produced the content, when, and on what basis? AI-generated content that reads as authoritative without attribution or disclosure is a framing risk in itself, particularly in regulated categories like health and finance.
How it works
You choose the publication channel, because the stakes scale the bar — a social post and a health-advice article are not held to the same standard. Then you work through twelve checks grouped into five areas: factual claims, currency, citations, framing, and accountability. Critical checks (verifying claims, checking stats and quotes, removing hallucinated citations) are weighted heavily. The tool computes a weighted score, applies a stricter threshold for high-stakes channels, and — crucially — blocks a “low risk” verdict if any critical step is still incomplete, listing exactly what to finish first.
Tips and notes
- Open every citation. The single highest-yield check: confirm the source exists and actually says what the text claims.
- Mind the cutoff. For anything recent — prices, laws, leaders, software versions — the model may be confidently out of date.
- Watch the framing, not just the facts. Cherry-picked true facts can still mislead. Check for false balance and implied causation.
- Human sign-off for high stakes. Health, legal, financial, and news content should clear this checklist and get qualified human review.