Do You Have to Tell Customers You Used AI?

“Should we tell people we used AI?” is really three questions wearing one coat, and most of the arguments I see are people answering different ones at each other. There is the legal question: what must be disclosed, where. The platform question: what the channels you publish on require. And the trust question: what disclosure does to the audience, which is where the research has turned up something genuinely uncomfortable. This is an opinion piece, so I will give you my answer up front: disclose synthetic media always, disclose chatbots always, and stop agonizing about labeling ordinary AI-assisted work - the label answers a question customers weren’t asking, in a format that damages trust without informing anyone.

Now the case for that position, one layer at a time.

What the law actually requires is narrower than you think

The legal duties are real but specific. In the EU, Article 50 of the AI Act has applied since August 2, 2026: people must be told when they are interacting with an AI system unless it is obvious, synthetic audio, image and video content must be marked in a machine-readable way, and deepfakes must be visibly disclosed. Notice the shape of that list. It targets impersonation of reality - fake humans, fake footage, fake conversation partners - not the use of AI as a drafting tool. There is even an editorial-review carve-out for AI-assisted text, which I walked through in my EU AI Act checklist.

The US has no general disclosure duty at all. California’s bot law, SB 1001, requires a bot that communicates with the intent to mislead about its artificial identity - to drive a sale or influence a vote - to disclose itself clearly and conspicuously. The FTC’s lever is deception, not disclosure: its fake reviews rule, in force since October 2024, bans reviews and testimonials from someone who does not exist, AI-generated fakes named explicitly. Nothing in it requires you to announce that AI helped write your product page - the full US picture is in my FTC guide.

So read the legal layer honestly and a pattern appears, and I find it clarifying: every binding rule is about not letting synthetic things pass as real ones. No regulator anywhere requires “this blog post was written with AI assistance.”

The platforms bind you faster than the law does

Where marketers actually get caught is platform policy, which ships quarterly and comes with enforcement built into the publish button. YouTube requires creators to disclose AI that meaningfully alters or generates photorealistic content - a real person appearing to say something they didn’t, altered footage of real events, realistic scenes that never occurred - while explicitly exempting fantasy content, beauty filters and production assistance like script drafting. Meta requires the disclosure-and-label step for photorealistic video or realistic-sounding audio that was digitally created or altered, and may penalize accounts that skip it. TikTok requires labels on realistic images, audio or video that AI generated or significantly edited.

Same pattern, third time: photorealistic, realistic-sounding, realistic-looking. The platforms converged on the law’s logic independently, because they are managing the same risk - audiences being shown fabricated reality. None of them asks whether a model helped you write the caption.

For a marketing team the operational consequence is simple, and it is the first half of my policy: anything synthetic that a viewer could mistake for a recording of reality gets labeled, on every channel, using each platform’s native tool. Not because a lawyer signed off on the risk, but because this is the one category where every rule-making body agrees, and the direction of travel is only one way.

The trust research is uncomfortable, and worth facing straight

Here is where it gets interesting, because the standard ethics-deck advice - “transparency builds trust, when in doubt disclose” - turns out to be empirically wrong in the general case. In thirteen experiments with over 5,000 participants, published in Organizational Behavior and Human Decision Processes, University of Arizona researchers found that disclosing AI use consistently reduced trust: students trusted a professor 16% less on learning AI helped grade, investors trusted firms 18% less when ads disclosed AI use, clients trusted graphic designers 20% less. The effect held across different ways of phrasing the disclosure, and it held even among people who use AI heavily themselves.

The same research carries the counterweight, and it is the load-bearing finding for me: being exposed by a third party is worse than disclosing. If someone else reveals your AI use - a detector, a journalist, a screenshot - the trust damage exceeds anything self-disclosure would have cost.

So the real decision matrix was never “disclose and build trust” versus “hide and risk it.” It is: disclosure has a known, measurable cost, and concealment has a larger cost multiplied by the odds of exposure. That reframing does most of my thinking for me. Where exposure is plausible - synthetic media, chatbots, anything a platform can auto-detect and label over your head - disclose, because you are buying out of the worst outcome at a discount. Where there is genuinely nothing to expose beyond “a professional used professional tools,” a label is all cost and no hedge. The same audience-sentiment math shapes the rest of the brand safety picture, where the research on how people receive visibly AI-made marketing lives.

The policy I actually run

Three tiers, and the whole thing fits on an index card.

Always disclosed, no judgment call: chatbots and anything conversational, labeled as automated at the start of the interaction. Synthetic or photorealistic generated media, labeled with the platform’s native tool and marked machine-readably where the AI Act requires. AI-generated voices. Anything where a reasonable person would otherwise believe they were seeing a recording of reality or talking to a human.

Disclosed as process, where clients or partners are owed it: in my consulting work, what I use models for and what never leaves human hands sits in the working agreement. That is disclosure at the relationship level, where the Arizona findings bite least - a specific explanation to a party who benefits, not a warning sticker for strangers. The same logic applies to bylined thought leadership: the byline is a claim of authorship, so the humans behind it should be able to describe their process without flinching.

Not labeled: AI-assisted research, drafting, editing, resizing, translation and every other tool-shaped use inside work a human reviewed and owns. My reasoning is not “you’ll get away with it.” It is that the label misinforms. A reader seeing “made with AI” on a carefully edited, fact-checked article receives the message “less care was taken here,” which in our case is false - the verification pass on AI-assisted work is more rigorous than most purely human workflows I have seen. A disclosure that creates a false impression is a strange servant of transparency.

The precondition for that third tier, and I want to be explicit about it: the work has to actually be reviewed and owned. Unreviewed model output shipped under a human byline is not an undisclosed tool, it is a misrepresented product, and no labeling policy fixes a process problem. That standard - every claim checked, every output edited, dark patterns kept out - is what makes the no-label tier defensible.

Where I think this lands in three years

A prediction, cheerfully falsifiable: process labels on ordinary content will disappear, and reality labels on synthetic media will become as unremarkable as the sponsored tag. The Arizona trust penalty exists, I suspect, because audiences currently read “AI” as a signal of corner-cutting. As AI-assisted becomes the default way professional content is made, the signal decays - you cannot infer corner-cutting from a practice everyone follows. What will not decay is the line every regulator and platform has already drawn: fabricated reality must announce itself. Build your policy on that line and you will not have to rebuild it each time a platform updates its rules.

FAQ

Do I legally have to disclose that a blog post was written with AI? In the markets I operate in, no - no US or EU rule requires it for editorially reviewed text. The EU AI Act’s text-disclosure duty targets AI-generated text published to inform the public on matters of public interest, with a carve-out where human editorial review takes place. Ads and product pages sit under ordinary deception law: the claims must be true, whoever drafted them.

Do AI chatbots have to identify themselves? Treat yes as the working answer everywhere. The EU requires it under Article 50 unless it is obvious, California requires it in commercial contexts on large platforms, and a customer who spends ten minutes believing the “agent” was human converts that discovery into exactly the third-party-exposure scenario the trust research says is most expensive.

Will disclosing AI use hurt my conversion rates? The honest reading of the evidence: a visible AI label on the work itself carries a measurable trust penalty today, on the order of 16 to 20% in the Arizona experiments. That is an argument for putting disclosure where it belongs - on synthetic media, on bots, in client relationships - rather than decorating every asset with it. It is not an argument for concealment where exposure is possible, because exposure costs more.

What should a small team write down? One page: the three tiers above, the platform label rules for the channels you actually use, and a named owner for the judgment calls. The point of writing it down is speed and consistency - the worst version of this is five marketers making five different calls under deadline. It is the same logic as the rest of the governance work: decide once, calmly, instead of repeatedly, in a hurry.

This post is part of the ethics and governance guides, where the legal layers get their own deep dives: the EU AI Act checklist and the FTC’s AI advertising rules. The wider practice starts at the complete guide to AI marketing.