Checklist
Check before you ship
Fluent AI output can still be wrong. A one-page checklist for checking it before it reaches anyone else, sized to the stakes. Print it and share it.
AI output reads well whether or not it’s right. A confident paragraph with a wrong number looks exactly like a confident paragraph with a right one.
Most people know they should check. Many don’t know how. Of AI users who used an output they had doubts about, 38.7% said they weren’t sure how to check it, and the most common reason for skipping the check was time (Charter AI Pulse, 2026). Executives are the most likely to use output they doubt.
First, name the stakes
Decide in advance which work always gets a full human check. For a People team, that list usually includes:
- Anything that affects a decision about a person: hiring, pay, performance, discipline, exits.
- Anything with policy, legal, or benefits content.
- Anything that leaves the team: leadership updates, all-employee messages, notes to candidates or vendors.
Everything else gets a quick pass. Put time on the calendar for the full check, or it won’t happen.
While you read
- Find the claims you’d repeat. Names, numbers, dates, quotes, policy references. Open the source yourself and find the exact line. A link that loads is not a source that agrees.
- Read it as the person it’s going to. An employee, a manager, a candidate, a lawyer. What would they question? Does it treat anyone unfairly? Does it include anything that shouldn’t leave the room?
- Ask for the case against it, in a fresh chat. Paste the output and ask what’s wrong, what’s missing, and what a skeptic would say. A new conversation hasn’t committed to the first answer.
- Check that it answers your question. Fluent output often answers a nearby, easier question. Reread what you asked.
Before you send
- Cut what you can’t vouch for. If you couldn’t confirm it, take it out, or say plainly that it’s unconfirmed.
- Get a second reader when the stakes are high. Not because you can’t be trusted. Because fluent text is easy to believe, including for you.
- Put your name on it. If you wouldn’t defend every line in a meeting, it isn’t ready.
Confidence isn’t evidence
In a recent Harvard survey, people were asked whether they recognized a list of names from the AI world that included two invented ones. Nearly half said they recognized the invented names (Shorenstein Center, 2026). Feeling sure isn’t being right, for people or for models. “Not sure, let me check” is a skill.
For leaders: say it out loud
When you use AI on something that matters, say how you checked it. “AI drafted this. I confirmed the numbers against the HRIS, and our benefits lead read the policy section.” It takes ten seconds, and it tells your team that checking is part of the job, not a sign of distrust.