Cheatsheet · Checking the work
AI fact-checking checklist
Models write fluent text, and some of it is false. Size your checking to the stakes. Check what would be costly if wrong, and skim what would not. These eight checks cover the claims that fail most often.
The checklist
Mark every specific claim
Underline each figure, date, name, quote, and citation in the draft.
Find the claims the decision rests on
Pick the claims that would change what you do if they were false. Check those first.
Open the source
Confirm each cited figure in the original document, with its context, before you use it.
Recompute the numbers
Re-derive totals, averages, and changes from the source data yourself.
Check quotes word for word
Compare each quotation with the original text, and confirm who said it.
Use an independent check
Use the source, fresh data, or your own calculation. Do not ask the same model if it is sure.
Recheck around a failure
When one claim turns out false, recheck every claim that leaned on it.
Cut, soften, or attribute the rest
If you cannot confirm a claim in time, remove it or state only what you can support.
Red flags
- A round figure with a tidy attribution and no source you can open.
- A study described as “widely cited” with no title, author, or date.
- A claim that “most experts agree” on something nobody could have measured.
- A citation you cannot find, which means the rest of the draft needs a second look.
At each band
- E1 · Foundational
- Reads everything before it ships and checks figures, names, and citations against a source.
- E2 · Proficient
- Checks the claims the decision rests on, using independent sources rather than the model.
- E3 · Distinguished
- Spots fabrication patterns, and recomputes or traces every figure headed to a client.
- E4 · Exceptional
- Builds checkpoints into the workflow where errors would compound, and sets the team’s standard.
Summarized from the Emeri study guide Checking the work, free at emeri.org/learn/checking-the-work.