How to disclose AI use, and where the disclosure belongs

The research says something uncomfortable

Most advice on this question assumes candor gets rewarded. The evidence points the other way.

Two management researchers ran 13 experiments with more than 3,000 people. The settings ran from classrooms and hiring to investing and creative work. In all of them, people who disclosed using AI were trusted less than people who said nothing. That penalty came from a judgment about legitimacy rather than a judgment about quality.

The researchers tested six different framings of the same disclosure. None of them removed the penalty. Whether the disclosure was voluntary or required made no difference either.

So disclosure carries a price. Anyone telling you it will win your audience over has not read the work.

The bigger penalty lands when somebody else does the telling

The same research tested what happens when a third party reveals the AI use. Trust fell further in that case than when the person disclosed it themselves.

That comparison is the whole decision. A standing cost you control sits on one side. A larger cost you do not control sits on the other.

Crisis communication has known this for a long time. Getting there first keeps you inside the story rather than outside it.

The exposure risk also depends on the work. Readers rarely detect the machinery, and what they detect is genericness.

Disclose the engine, never the author

Here is the distinction that makes the rest of this easy. Your disclosure describes how the work gets made. It does not hand over authorship.

The thinking is yours. The claims are yours. The judgment about what ships and what gets cut is yours.

Write the disclosure so a reader finishes it knowing who is accountable for the sentences. A disclosure that leaves that unclear invites the exact doubt it was meant to prevent.

Where the disclosure belongs

Put it in one permanent place a reader can find. A method page, an about page, or a short colophon works. One paragraph is enough.

Skip the badge on every article. A label on each piece attaches the trust discount to each piece. It also says nothing the standing page does not say better.

The person who wants to know how your work is made will go looking. Answer that person completely, in one place, and let the work speak everywhere else.

One exception holds. Where a platform or a regulator requires a per-item label, follow the rule as written.

What the disclosure should say

Four things belong in it. The mechanism, in plain words. What the machine does. What you do. Who answers for the result.

Something like this, adapted to your own arrangement:

The articles here start as recorded interviews with me, transcribed and turned into a set of guides that a Claude project works from. The project drafts. I direct it, cut what is wrong, and add what only I know. Every claim on this site is mine, and I answer for it.

Four sentences, no apology, no hedging. Nothing in there asks the reader to be comfortable with anything.

Skip the tone that treats the disclosure as a confession. Where your judgment enters and where the machine stops is worth its own answer. A reader who wants that detail is a reader worth writing it for.

Where a label is required, and where it is not

Two rules matter to most expert businesses right now.

The EU AI Act's transparency obligations took effect on 2 August 2026. They cover four situations. Telling people they are talking to an AI. Marking synthetic content in machine-readable form. Informing people about emotion recognition. Disclosing deepfakes, and AI-generated text on matters of public interest. Ordinary marketing writing does not sit in that list. The machine-readable marking duty falls mainly on the companies that build the tools.

Amazon's publishing rules run on a distinction worth borrowing. Amazon asks you to disclose AI-generated text, images, and translations when you publish a book. Work you wrote yourself and then improved with AI counts as AI-assisted, and that needs no disclosure. The answer goes to Amazon rather than onto your product page.

These rules move. Check the platform you publish on before you publish. Do not trust a summary from last year, including this one.

The rule with no exceptions

Never publish a claim you cannot defend in the other direction. A "written entirely by a human" badge on a page that had help is the exposure case. It waits for somebody with a screenshot.

The reverse fails the same way. Claiming a machine did work you did by hand, to sound current, is the same move. Both cost more than the thing being covered up.

What disclosure will not do

Disclosure will not make weak work acceptable. A reader who finds an article thin does not soften on learning how it was assembled.

It also will not settle an accusation on its own. Someone convinced your writing came from a machine wants evidence of your thinking. A policy statement is not evidence. The work has to carry it.

And it is not a marketing asset. A campaign built on your own transparency draws the discount described above and buys none of the protection. The protection comes from having answered before anybody asked.

What you are buying

A disclosure page will not earn applause. Readers do not thank anybody for it. The research says a few will think slightly less of you for reading it.

What it buys is the sentence you would rather write yourself than have somebody else write about you.

The people who go looking will find it. They will also find that you told them first, in your own words, before anyone asked you to.