What to say when a client asks whether AI wrote it

Do not lead with a detector score

The instinctive move is to run the piece through a detector that gives a friendlier number and send that back. It is the worst available option, for a reason that is not obvious in the moment.

Sending a competing score concedes the premise — that a detector is the arbiter of whether your work is legitimate. You have now agreed to be judged by a tool you do not control, which changes its model without telling you, and which the client can re-run tomorrow with a different result. You will win this exchange and lose the next one.

It is also standing on ground that will not hold. Detection tools disagree with each other on the same text, and they produce false positives on genuine human writing — with a bias against writing by non-native English speakers that Stanford researchers documented, and against any prose that is plain, structured and unadorned. Which is to say: against good business writing.

The answer is to move the conversation off the score and onto evidence you actually have.

What they are really asking

Almost nobody sends that email because they care about methodology. Read it again and it is usually one of these:

• "I think I am being overcharged." The detector gave them a number to hang a suspicion on.
• "My boss asked me and I do not have an answer." They need something they can forward. This is the most common one, and the easiest to solve.
• "I do not think this is very good." The AI question is a proxy for a quality complaint they cannot articulate. Answering only the AI question leaves the real problem in place.
• "I am worried about being penalised by Google." A genuine, answerable concern.

Your reply should work for all four, which means it needs to cover process, accountability and quality — not just deny the accusation.

The evidence that actually holds up

Detector scores are contestable. These are not:

• Version history. A document with two hundred revisions over three days, showing sentences being reworked, is close to unanswerable. Google Docs keeps this automatically. If your team drafts in Docs, you already have your best evidence and did not have to do anything.
• The brief and the research trail. The notes, the sources, the interview recording, the questions you asked their product team. Generated content does not have a research trail behind it.
• Specifics only a human could have. The detail from the site visit, the thing the sales manager said on the call, the number from their internal dashboard. Point at three of them in the piece.
• The writer. Offer a fifteen-minute call with the person who wrote it. Nobody who generated an article can talk about why they structured it that way for ten minutes.
• Your stated policy. If you have already published how you work, you are pointing at an existing document rather than improvising a defence. That difference is doing a lot of work.

Notice that four of the five are things you either have or do not have before the email arrives. The response to this problem is mostly preparation, not rhetoric.

The reply

Calm, specific, no defensiveness, and it ends by offering more rather than less. Adapt the bracketed parts — sending this verbatim will read as a form letter, which is its own kind of tell.

Subject: Re: [article] — happy to walk you through how it was made

Hi [name],

Thanks for flagging it — I would rather you asked than wondered.

The short answer: [writer's name] wrote this piece, and [he/she/they] is
accountable for it. Here is what sits behind it:

- The brief and outline we agreed on [date]
- [Interview / call / site visit] with [person] on [date], which is
  where the detail about [specific thing] in the third section comes
  from
- [N] sources, all linked in the draft
- The full editing history, which I can share if it is useful

On the checker itself: those tools are unreliable in both directions.
They disagree with each other on identical text and they flag genuine
human writing regularly — particularly clear, structured business prose,
which is what we are deliberately writing for you. I could send you a
different tool's score showing 2%, but I do not think either number
tells you anything worth having.

What I would rather offer: [writer] is happy to spend fifteen minutes on
a call walking you through why the piece is built the way it is. That is
a much better test than any detector.

One thing worth me being straight about — we do use AI tools in our
process, for [research summaries / outlines / alternative phrasings].
Every deliverable is written and verified by a person before it reaches
you, and we never let a tool near a fact, figure or quote without
checking it against the source. Our full policy is here: [link]

If the underlying worry is that the piece is not landing the way you
hoped, tell me that instead and we will fix it — that is a much more
useful conversation than the detector one.

[name]

The final paragraph is the most valuable one. In a good proportion of these cases, the real complaint is quality, and inviting it directly turns an accusation into a revision request.

If it was flagged and AI was involved

Sometimes the honest answer is that a model drafted it and the editing was lighter than it should have been. Two rules here.

Do not lie. Not for ethics alone — for arithmetic. A denial that later collapses costs the account and the referrals. An admission costs one uncomfortable email.

Reply with a fix, not an apology. "You are right that this one leaned on a draft more heavily than our standard. Here is what we are doing about it, here is the rewritten piece by Thursday, and here is the process change so it does not recur." Clients forgive process failures that come with a correction attached far more readily than most agencies expect.

Then actually make the process change, because the second occurrence is not survivable.

Making the question stop arriving

The agencies that get this email rarely are not the ones producing purer content. They are the ones who made the process visible before anyone asked.

Three things move the needle: publish how you work, so the answer exists in advance. Name the writer on the delivery, because attribution makes accountability visible. And attach something to the deliverable — a short note on how the piece was made, what it was checked against, and who approved it.

That last one is what Wordcheck produces: a report under your own branding showing how the writing reads, what it was compared against for originality, and every score explained in a sentence the client can follow. It ships with the invoice, and the question tends not to come up, because it has already been answered. More on that on the agencies page.

Common questions