Using AI honestly: a practical ethics check
A short check to run before the work leaves your desk.
Most "AI ethics" content is either too abstract to use or too cautious to be useful. This isn't either of those. This is a short, practical check you can run on any piece of AI-assisted work before you ship it.
Six questions. Each one is concrete. If you can answer all six in a way you're comfortable with, you've covered the basics. If you can't, the question itself tells you what to fix.
The two that catch the most
Did I check it? AI output can be wrong, misleading or partly invented, and reading it is not the same as verifying it. Every claim, name, number and source needs to hold up. If you didn't check, what you have isn't AI-assisted work. It's AI-substituted work with no quality control.
Did I add real value, or did I just pass it through? There is a difference between using AI inside your work and presenting its output as your thinking. The first is fine: you drafted, restructured, summarised, accelerated, and your judgement is visible in the result. The second means the client is getting AI output with a few words changed. The test is whether you can point at specific places where your knowledge changed the work.
The six questions
Did I check it?
Did I add real value, or just pass it through?
Would I be comfortable if the people involved knew?
Did I use someone else's information without consent?
Am I representing this honestly?
Could this cause harm if it's wrong?
If the answer is no
Go and check it now
Add real value, or rewrite it
Disclose it, or change the approach
Anonymise, get consent, or stop
Describe the work accurately
Verify in proportion to the harm
Ethics, in practice, is mostly small honest decisions made quickly.
The three about other people
Would I be comfortable if they knew? The client, the colleague, the audience, the person whose data went into the tool. This is the gut check that catches most of the real problems, and it's worth being precise about what it tells you. If you'd be embarrassed to disclose the AI use, the problem isn't the disclosure. It's the underlying decision.
Did I use someone else's information without consent? Client data, personnel details, supplier contracts, partner conversations. If you don't know whether you had the right to put it in, you probably didn't have explicit consent.
Am I representing this honestly? This one is about how you describe the work rather than the work itself. Billing for time you didn't spend, implying original thought that wasn't there, claiming expertise you borrowed. If someone asked directly, your answer shouldn't surprise them.
Scale the checking to the damage
The last question is what happens if this turns out to be wrong. For a draft email or an internal summary the answer is embarrassment and an easy fix. For legal wording, medical content, financial guidance, or anything that shapes a decision about someone's livelihood, the answer is real damage.
So the verification should scale to it. Low stakes, one careful read. High stakes, expert review, more than one check, and slow down.
Three things to do this week
01
Run the six questions before it leaves your desk
It takes under a minute and catches nearly everything worth catching.
02
Treat an uncomfortable answer as information
The point isn't to feel bad about it. It's to know which specific part needs fixing before it ships.
03
Fix the process, not just this one piece
If the same question keeps failing, the workflow is wrong, not that single deliverable.
The six-question check
Run these before you ship any AI-assisted work. 1. Did I check it? Every claim, name, number and source verified. 2. Did I add real value, or did I just pass it through? I can point to where my judgement changed the work. 3. Would I be comfortable if the people involved knew? The client, the colleague, the audience, the person whose data went in. 4. Did I use someone else's information without consent? The data was mine, covered by an agreement, or fully anonymised. 5. Am I representing this honestly? My description of the work matches how it was produced. 6. Could this cause harm if it's wrong? Verification scaled to the stakes.
Conclusion
01
Six questions: did I check it, did I add real value, would I be comfortable if they knew?
02
Did I use someone else's information without consent, am I representing this honestly, could it cause harm if it's wrong?
03
If you can't answer one of them well, that question is telling you exactly what to fix.
Go to next