Checks one and two: specifics and sources

Check one: every specific. Go through the output and mark every date, number, name, quotation and reference. For each, either verify it against a source you trust, or remove it, or replace it with a hedge ("in the region of," "reportedly"). No specific survives unchecked into anything with your name on it.

Check two: every source. If the output cites anything, find it. Not a similar thing; that thing. Open the link. Search the title. Check the author and year. If you cannot find it in two minutes, it does not exist for your purposes, and the claim it supports goes back to unverified. Module 03 covers how to get better sources in the first place.

These two checks catch the majority of harmful errors, because harm mostly comes from a specific claim that turns out to be wrong.

In practice. An analyst ran checks one and two on a two-page AI-drafted briefing. Fourteen specifics. Nine verified. Three were wrong and were corrected. Two could not be verified either way and were rewritten as hedged statements. Of three cited sources, one did not exist. Total time: eleven minutes. The briefing went to a director.

At your desk. Take a real piece of AI output you intend to use. Do check one: mark every specific, verify or remove each. Do check two: find every source. Record the counts: specifics found, verified, wrong, removed; sources found, real, missing.

Write it down.

• Specifics: found
• verified
• wrong:
• Sources: cited
• real:
• Minutes it took: