Sounds Good ≠ Correct — Own the Output

Here’s the point, stated plainly: working with AI requires the same work attitude and work ethic as any other work output. Nothing has changed.

If a junior colleague handed you a draft, you wouldn’t forward it because it “sounded finished.” You’d check the claims, fix what you couldn’t stand behind, and own the send. If you wrote the email yourself, you’d do the same. An AI-assisted draft gets that same standard — not a special AI ritual, not a lower bar because the prose looked polished, and not a higher bar that invents a new profession. The tool changed. The bar didn’t.

AI drafts often sound finished — headings, bullets, a decisive recommendation — so you almost forward them. That’s the trap: text that feels complete can still carry unsupported claims, wrong names, invented metrics, and missing caveats. If your only filter is “does this sound professional?”, you will ship style and call it substance.

You don’t need a perfect evaluation framework. You need the same diligence you’d apply before sending any work product — one named check, every time, so “sounds good” never becomes “sent” on autopilot.

In One Useful Task, verification sits inside the workflow — the same place it always belonged — not as a disclaimer you remember after someone spots the error in a meeting.

Confidence is a style. Correctness is ownership.

Try it safely

Use a deliberately flawed sample so you practise catching errors without risking real work.

Synthetic flawed response (imagine AI wrote this from a short email thread):

“Q3 churn dropped 18% after we launched the new onboarding wizard on 12 March. Finance confirmed the payback period is 4.2 months. Recommend rolling the wizard to all regions next Monday; legal has no concerns.”

What’s wrong with a paragraph like that? Possibly everything: the percentage, the date, whether Finance “confirmed,” whether legal was asked, whether “all regions next Monday” was ever on the table. The problem isn’t that AI is always wrong. It’s that wrong and right can look identical until you check — exactly as they can in a rushed human draft.

You wouldn’t stamp that paragraph “good enough” because it sounded decisive. You don’t stamp an AI draft that way either.

The method — pick ONE default check

Before you send, paste, or present anything AI drafted this week, choose one primary check and stick to it until it’s automatic. This isn’t an “AI tip.” It’s the same pre-send diligence you’d apply to any work product that carries your name.

Check A — Source compare
Open the original email, notes, or doc. Highlight every factual claim in the AI draft. Can you point to the line that supports it? If not, flag it.

Check B — Invention scan
Ask: “What in this draft was not in my inputs?” Names, dates, metrics, motives, quotes, “Finance confirmed…” — mark each invention. Delete or demote to a question.

Check C — “What would make this wrong?”
Write one sentence: the single fact that, if false, collapses the recommendation. Then look for that fact on purpose.

Optional micro-prompt to run after you have a draft (still verify yourself) in your AI chat tool:

Review the draft below against the source material.
List:
1. Claims supported by the source (quote the supporting line)
2. Claims not found in the source (possible inventions or assumptions)
3. Missing caveats or unresolved questions
4. What would make the main recommendation wrong

Do not improve the prose yet. Only audit.

Source:
[paste source]

Draft:
[paste draft]

Use the audit as a mirror — not as a second authority you trust blindly. AI can mis-audit too. Your eyes still own the send button — the same way they would if a colleague wrote the first pass.

What happened when you practise this

People who treat AI output with ordinary professional ownership usually notice the same pattern:

• The first draft looks “done.”
• The check finds 2–5 items that weren’t in the inputs.
• Fixing those takes less time than recovering from a wrong forward.
• The habit transfers: you start briefing better (Article 4) because you hate cleaning invented detail.

That’s the point of a Work Recipe: not a cleverer prompt, a repeatable way of working — good work, with AI in the loop.

Check before you use it (meta-check)

• You named which check you used (A, B, or C)
• Consequential claims are source-linked or labelled uncertain
• Recommendations that need human approval are not phrased as done deals
• You did not treat “the model said it was fine” as verification
• You would stand behind this the same way you’d stand behind a junior’s draft or your own email

Essential lesson: Confidence is a style. Correctness is ownership.

Make it yours / when not to rely on one check alone

Adapt: For emails, Check B (invention scan) is often enough. For numbers and dates, Check A. For recommendations that commit other people, Check C plus a human confirm.

Don’t stop at one check when: Legal, financial, medical, HR, or customer-facing commitments are involved; the draft will be forwarded beyond people who know the context; or you couldn’t provide source material in the first place (then AI is brainstorming, not reporting — label it that way). High-stakes work never got a free pass before AI. It doesn’t get one now.

Your challenge

Before anything AI-assisted leaves your desk this week, run one named check — with the same seriousness you’d give any other send. Then note:

• What did the check catch?
• What would you have sent without it?

Did this produce something you could use, and what did you have to correct?

Tell us which check you chose. Next: Before you paste — what shouldn’t go into the chat. New here? /start-here · /brief