AI Fact-Check Checklist

How is the verification plan prioritised?

Checking everything equally wastes effort on claims that carry no risk. The checklist sorts claims by consequence: numbers and statistics, named people and organisations, dates and chronology, quotations, legal or medical statements, and finally general background.

Each category gets a verification method rather than a vague instruction. Statistics need a primary source, not a secondary article citing one. Quotes need the original recording or transcript. Legal and medical statements need the current text of the rule, since these change and models are trained on snapshots.

The checklist finishes with the two failure modes that survive most reviews: outdated facts that were true when the model was trained, and confident specifics on subjects too niche to have reliable coverage.

How do you use the AI Fact-Check Checklist?

1. 1.Highlight every number, name, date and quotation in the draft before checking anything.
2. 2.Verify each against a primary source and record the link next to the claim.
3. 3.Check publication dates on sources; a correct fact from 2022 may be wrong now.
4. 4.Delete anything you cannot source rather than softening it with 'reportedly'.

What can this tool not tell you?

• It is a process, not a database — it cannot verify claims for you.
• Primary sources are not always public, particularly for commercial data.
• Thorough checking is slow; budget for it rather than assuming it is a final skim.

Why not all claims deserve the same scrutiny?

Treating every sentence in a draft as equally worth checking guarantees that fact-checking either takes far longer than the piece warrants or gets abandoned halfway through, and either outcome leaves the highest-risk claims — the ones that could actually damage credibility if wrong — no better verified than the low-risk ones. Prioritising by consequence rather than checking sequentially from the top of the document is what makes thorough verification achievable within a normal production schedule.

Numbers and named entities sit at the top of the priority list because they're both high-consequence and easy to get definitively wrong: a statistic can be misremembered by an order of magnitude, a company name can be confused with a similarly-named competitor, and neither error looks wrong on the page the way a garbled sentence would. Quotations are similarly dangerous because a paraphrase that drifts slightly from the original, presented as a direct quote, is a factual and often reputational error even when the underlying meaning survives.

The two failure modes that persist even after conscientious checking — outdated facts and confident coverage of niche subjects — deserve specific attention because they don't feel like risks while writing. A fact that was true eighteen months ago doesn't announce itself as stale, and a model writing fluently about a small, under-documented subject doesn't sound any less certain than when it's writing about something well established, which is exactly why source dates and topic obscurity need to be checked deliberately rather than assumed away.

Recording the source next to each claim, rather than just confirming it and moving on, pays off later in ways that aren't obvious during the first check. If an editor, a legal reviewer or a reader later challenges a figure, the writer who kept a running list of sources next to each claim can respond in minutes; the writer who verified everything mentally and moved on has to redo the entire search from scratch under time pressure. This habit also surfaces a second-order benefit: claims that turn out hard to source at all, because no primary reference actually exists, get caught during the writing process instead of after publication, when the cost of walking them back is much higher.

What do worked examples look like?

What do people ask most about this tool?

Which related tools should you try next?