Run the AI's answer through three checks first

A citation can look complete and still be a dead link. A single source can sound certain and still not count as a cross-check.

Monday morning, before standup, you ask AI whether a new policy will hit the team. It gives a clear conclusion, plus an “industry association report,” a “research-institute brief,” and three links. The page is laid out like a memo that's already been researched.

You don't have time to open them one by one, so you paste the conclusion straight into the briefing. When the manager asks for evidence, the first link is a 404, the second dumps you on the homepage, and the third opens — but the content is the same association report. Three links. One piece of evidence.

The crash is specific: AI was both the answerer and the checker. It can dress up what it just wrote so it looks more like evidence, but repeating itself is not an outside check. You need to give verification three different jobs.

Alice

Online

AI just handed you a claim ready for the briefing: “The new policy will cut filing costs for small and mid-size teams by 30%.” Which check instruction do you send first?

Label a source on every claim, and paste the supporting sentence and the link. Recheck this conclusion from another angle, and list where the two answers disagree. Confirm the last answer is fine, and explain why these claims are worth believing.

Send to Alice

Three gates, three different problems:
Ask for sources: Make every claim point back to an original you can open. Stops content invented from thin air.
Cross-check: Find independent sources and check them against each other. Stops bias from a single piece of evidence.
Flag uncertainty: Separate what's confirmed, what's inferred, and what's missing. Stops unknowns being written as facts.

Now set up a full check. Pick 3–5 actions below, and I'll tell you how far this set of checks can take you.

Require a quote from the original for each claim; write “not found” if it isn't there Open each link and match the page title to the quoted sentence Find material from a separate organization and cross-check List conflicts between sources on their own Mark what's confirmed, inferred, and still missing Have AI score how trustworthy its own answer is Have the same AI answer twice in a row and see if it matches Have AI rewrite the conclusion so it sounds more certain and more professional

Submit check actions 0 selected, need at least 3 more

Evidence

Try another set of actions

Don't write the check result as just “trustworthy” or “not.” A result you can act on needs at least four blocks: the judgment, the evidence, the next step, and open questions. Whoever picks it up later knows what they can use, and what's only a lead.

Even if all three pass, open the most important original. Cross-check can shrink the error range; it can't tell you what the page actually says. The model shrinks what still needs a check. You make the key evidence land.