Using AI without hiding it

1. Name the work and the check

   Name the part AI touched and the source you checked, following the lab's editorial standards. Here's the part you can copy:

   Drafted with AI, checked against [source] by me.

   Copy disclosure

   Replace “drafted” with the work AI actually did, and fill in the source you opened.

   You know it worked when your line names the AI's part and your actual check, as the site's authorship disclosure does.

2. Keep the source beside the draft

   Attach the exact passage or record you used, the way the editorial standard ties each account to the operating record. Mark anything you cannot verify as unverified, or leave it out of the draft.

   You know it worked when another reader can follow a claim back to its source, applying “What reviewed means”. Identify internal evidence without making private files public.

3. Run a check that isn't you

   Give a colleague the source and ask them to challenge the claim, or use a second tool to check the result. For a model review, follow the lab's rival-model sign-off account: use an out-of-family reviewer, a model from a different family than the one that did the work. Write what must pass before the review starts.

   Make that review a gate, a checkpoint that can stop the work, for the high-stakes result you are checking. Hold it when the review fails and reproduce disputed findings, as the lab did with the rival model's false build alarms.

   You know it worked when you can verify the reviewer's evidence and hold or clear the work, following the held release and repeated build checks.

4. Say what you decided

   Write your decision beside the checked result, using the lab's money-page sign-off as the pattern. Name what stays with the owner, the decisions AI cannot make for you. Keep approval of a changed offer with the person responsible for it, even when the draft passes its checks.

   You know it worked when your approval or hold is on record, just as a lending money-page change waits for Nate's sign-off.

5. Keep a record you could show

   Keep a receipt, evidence from the actual result that shows whether your check passed. Save the version you checked with its source. Add the check result and who ran it, using the lab's claim-to-receipt chain as your pattern.

   You know it worked when someone can repeat your check on the delivered work, the standard behind the rule against hollow “done” reports.

6. Disclose once where the work lives

   Put the line and its source beside the result you hand over, following the lab's on-site disclosure. Link the receipt there so the reader can inspect the check. Put any later correction in that same place with its date, as the editorial correction rule requires.

   You know it worked when the recipient can find the disclosure with the work, and a correction stays on the affected page under the lab's published standard.

• Calling a status report proof

  Check the result before repeating “done.” In the hollow “done” account, a blank page was called complete and an unwired safety mechanism was called armed. Require evidence another person can check before you pass that claim on.

• Trusting a checker that never ran

  Trace a real result through the check you name in your disclosure. The lie-detector that wasn't plugged in passed its own tests but never saw the live status reports. Keep “built and tested” separate from “ran on this work.”

• Accepting the confident wrong draft

  Open the cited source and find the passage supporting the claim. Use the failure modes page's examples as your warning signs: a reference that cannot be found, or a summary that adds a claim the source never made. Remove the unsupported claim even if the draft reads cleanly.

• Calling a green build a checked page

  Open the thing a person would see before you call it checked. In the 463-tests, blank-website account, a security rule left the browser showing nothing while every test passed. Compare the actual rendering, as the lab did: zero characters with the broken rule, 1,763 with it off.

Fictional practice, separate from the lab accounts above.

Use the lab's owner-approval boundary on this fictional reply: a customer wants a refund for a late order, but you have authorized only a free express replacement. The AI draft says, “I've issued your full refund and added a discount.”

Compare the draft with the authorized offer, then remove the invented refund and discount. Write your disclosure beside the corrected reply and attach the authorized offer as its source. Keep the decision about any refund with you, following the client delivery owner's approval rule.

Pick the last thing you used AI on and write its one-line disclosure using the template above. Attach the source you checked it against and put both where that work lives. Use the editorial standard to keep the line honest: leave “checked” out until you have opened the source and compared the claim.

You know it worked when a colleague can read your line and inspect the source without asking you to reconstruct the check, meeting the lab's checkable-evidence rule.