- Three separate regimes start biting around August: the EU AI Act's Article 50 (from August 2), New York's synthetic-performer law (since June 9), and Amazon's marketplace policy. They cover different things. [1][3][5]
- The EU forces machine-readable marking plus visible deepfake disclosure, with penalties cited up to 15M euros or 3% of turnover. [2]
- New York targets fake humans in ads ($1,000, then $5,000 per violation). Amazon targets photorealistic AI-generated people in listings, not all AI imagery. [3][5]
- The claim that an AI label tanks conversion has no primary-source field study behind it. It is a contested commercial assumption. [7]
- Do the operator move: split creative by trust sensitivity (human for endorsement assets, AI for low-stakes), and test labeled against unlabeled on your own traffic.
For two years the compliance conversation about generative creative was theoretical. On August 2 it stops being theoretical in the European Union, and it has already stopped in New York and inside Amazon. If you run paid social, marketplace listings, or a website that serves European or New York audiences, some of the creative you shipped last quarter now needs a label, and some of your competitors' does too.
The reflex reaction in every agency channel is the same worry: the label will scare shoppers off and cost you conversions. That worry is repeated so often it has the texture of a fact. It is not one. The regulation is real, dated, and enforceable. The conversion penalty is a rumor that has never been measured cleanly. Sorting those two apart is the whole job, because if you treat an unproven fear as a planning input you will make expensive creative decisions to avoid a cost that may not exist.
01Three regimes, three different definitions of "AI content"
The first mistake is to talk about "the AI label" as if it were one rule. It is three, and they do not agree on what they are labeling.
| Regime | What it actually covers | The teeth |
|---|---|---|
| EU AI Act, Article 50 [1] | Machine-readable marking of AI outputs, plus visible disclosure of deepfakes and some AI-generated public-interest text. Applies from August 2, 2026. [2] | Penalties cited in legal commentary up to 15M euros or 3% of worldwide turnover [2] |
| New York, synthetic-performer law [3] | A conspicuous disclosure when an ad contains a "synthetic performer" (a digital asset impersonating a human) and the advertiser knows it. Effective June 9, 2026. | $1,000 first violation, $5,000 each after [3][4] |
| Amazon marketplace [5] | Sellers must tag listing images and videos containing photorealistic AI-generated people; a shopper-facing indicator may appear. [6] | Listing suppression and account risk, per Amazon policy |
Read down the middle column and the shape of the problem appears. The EU cares about whether a machine can detect that content is synthetic, and separately about whether a human viewer is warned when they see a deepfake. New York cares about one specific thing: a fake person standing in for a real performer. Amazon cares about an even narrower thing: photorealistic AI-generated people in a product listing. None of them says "label everything you made with AI." An AI-generated background, a product render, a real photo retouched with AI tools: mostly outside all three, mostly the moment you assume otherwise you start slapping disclosures on assets no law asked you to, teaching your own shoppers to distrust creative that was never the concern.
There is one timing nuance worth holding. The EU's transparency duties apply from August 2, but the marking and watermarking obligation for systems already on the market before that date is delayed to December 2, 2026. [2] That is a few months of runway on the machine-readable side, not on the human-facing deepfake disclosure.
02The conversion panic that nobody has measured
Now the part everyone gets backwards. Search for whether AI labels hurt performance and you will find confident numbers: this many percent drop in click-through, that many percent loss in trust. Follow the numbers back and the trail goes cold. The most-quoted figures come from social posts and secondary summaries, not from a controlled field test on real buyers.
That is not a claim that labels are harmless. It is a claim that the honest state of the evidence is "we do not know," and the more careful reading actually points somewhere more useful. What seems to depress response is not disclosure itself but perceived artificiality: low-effort, obviously synthetic creative that a shopper clocks as cheap. And how you word the disclosure appears to matter more than whether you include one. In one study circulated through industry summaries, a "Human-made, AI-enhanced" framing scored better on trust than a flat "AI-generated" label. [7] Same underlying asset, different sentence, different response. If that holds, the lever is not "use less AI." It is "be precise about what the human still did."
So the operator question is not the panicked one ("will the label cost me sales?"). It is two sharper ones: which assets actually carry a trust burden, and does a disclosure move the number on your traffic. The first you can reason about today. The second you can test.
03Split creative by trust, not by cost
The durable way to think about this is creative architecture: decide where a synthetic asset is fine and where a human still earns the premium, before compliance forces the question in a panic. The dividing line is trust sensitivity. Some creative is a proxy for identity, endorsement, or authenticity, and a disclosed synthetic human quietly undercuts the exact thing that asset was doing. Other creative carries no such burden, and AI is simply the cheaper, faster way to make it.
Keep human (high trust burden)
- Hero images and video where a face carries the endorsement.
- Founder, team, and customer-testimonial content. The whole value is that the person is real.
- Pricing, claims, and high-consideration pages, where a labeled synthetic presenter reads as a tell.
Use AI freely (low trust burden)
- Backgrounds, textures, and lifestyle scenes with no synthetic person in them.
- Variant generation: sizes, crops, seasonal restyles of an approved asset.
- Internal and mid-production stages that never reach the consumer as the final, disclosed frame.
This split does two things at once. It concentrates your human production budget where it defends a real signal, and it keeps most of your AI volume outside the regimes entirely, because an image with no synthetic person is not what New York or Amazon is regulating in the first place. You comply by design instead of by disclaimer.
04Then test the label, do not fear it
Where a disclosure is genuinely optional (outside the mandatory cases, on your own site, in your own paid social), stop guessing and run the test the commentary never did. It is a plain creative experiment, and the instinct is the same one behind every real test that beats a dashboard.
- Take one real, converting asset. Run it in two cohorts: with the disclosure and without, everything else held equal.
- Measure the number that pays the bills, conversion or cost per acquisition, not just CTR. A label can move the click and not the sale.
- Test the wording, not only the presence. Put "Human-made, AI-enhanced" against a flat "AI-generated" against no label, since framing looks like the bigger lever. [7]
- Hold it long enough to clear noise. A few hundred conversions per arm, not a two-day peek.
- Read it per asset type. The effect on a faceless lifestyle shot will not be the effect on a testimonial, and averaging them hides the only finding that matters.
Run that once and you replace a secondhand rumor with a first-party number, which is the only number you should let change a budget. Most brands will find the effect is small, uneven, and swamped by whether the creative is any good, which is exactly what the honest reading of the evidence predicts, and the same pattern as AI's real but modest effect on conversion once you measure it instead of repeating it.
The regulation is not the interesting part of this story, and it is certainly not optional: mark what the EU says to mark, disclose the synthetic humans New York and Amazon are asking about, and move on. The interesting part is that a whole industry is about to make cautious, expensive creative decisions to dodge a penalty nobody has shown exists. Label because the law says so. Keep humans where trust is the product. And decide the conversion question the way you decide every other one that matters, by testing it, not by repeating it.
Sources
- European Commission · Code of Practice on Transparency of AI-Generated ContentArticle 50 transparency obligations apply from August 2, 2026; machine-readable marking plus visible deepfake disclosure
- Stibbe · The AI Act's transparency obligations: rules, scope and timelinethe August 2 application date, the December 2 marking delay for pre-existing systems, and the penalty ceilings cited for Article 50 breaches
- McDermott Will & Emery · New York's synthetic-performer disclosure law: what advertisers need to knoweffective June 9, 2026; conspicuous disclosure for synthetic performers; $1,000 then $5,000 penalties; audio and translation exemptions
- Transparency Coalition · Using AI in an ad? You've got to disclose it, starting this month in New Yorkplain-language summary of the New York disclosure requirement and its scope
- CNBC · Amazon makes sellers label AI-generated people in images after New York lawAmazon's requirement is scoped to photorealistic AI-generated people in listings, with a shopper-facing indicator, not all AI imagery
- Amazon Advertising · AI-generated content disclosure guidancethe seller-side metadata tagging workflow for AI-generated people in creative
- PPC Land · Advertisers face $5,000 New York fines and 3% EU penalties over AI labelsroundup of the label regimes and the state of the disputed, largely uncited claims that AI disclosure depresses response; framing appears to matter more than presence
Frequently asked questions
When do AI creative disclosure rules take effect in 2026?
The EU AI Act's Article 50 transparency obligations apply from August 2, 2026, though a marking and watermarking deadline for systems already on the market before that date is delayed to December 2, 2026. New York's synthetic-performer disclosure law took effect June 9, 2026. Amazon began requiring sellers to tag listing images and videos containing photorealistic AI-generated people in July 2026. They are three separate regimes with three different scopes, not one label.
Does a 'made by AI' label actually reduce click-through or conversion?
There is no primary-source field study proving a general penalty. The claim is repeated across marketing commentary but the numbers rarely trace to a controlled experiment. Some research points the other way: how the label is framed appears to matter more than whether it exists, and low-effort or obviously artificial creative depresses response more reliably than an honest disclosure does. Treat the conversion effect as an open question to test on your own traffic, not a settled fact to plan around.
What does the EU AI Act Article 50 require for AI-generated content?
Providers of generative AI systems must mark outputs in a machine-readable format so they are detectable as artificially generated or manipulated, and deployers must visibly disclose deepfakes and certain AI-generated public-interest text. It distinguishes machine-readable marking, visible disclosure, and deepfake labeling as separate duties. Breaches can draw penalties cited in legal commentary at up to 15 million euros or 3% of worldwide annual turnover, so it is not a soft obligation.
What does New York's synthetic-performer law cover?
It requires a conspicuous disclosure when an advertisement contains a synthetic performer, a digital asset meant to impersonate a human performer, and the advertiser has actual knowledge of it. It carries civil penalties of $1,000 for a first violation and $5,000 for each later one, with exemptions described for audio-only ads and translation-only uses. It targets fake humans specifically, not every AI-assisted image.
Does Amazon's policy apply to all AI-generated images?
No. Amazon's requirement is narrow: it applies to listing images and videos that contain photorealistic AI-generated people, and Amazon says it will add a shopper-facing indicator where applicable. AI-generated backgrounds, product renders, and real photos edited with AI tools are not the target. Reading it as 'label all AI content' will make you over-disclose assets the policy never meant to catch.
How should I decide which creative to make with AI versus a human?
Split by trust sensitivity, not by cost. Reserve human creative for assets that act as a proxy for identity, endorsement, or trust: hero images with faces, founder and testimonial content, pricing and high-consideration pages. Use AI more freely where the shopper does not rely on human authenticity as the value signal: backgrounds, variants, lifestyle scenes without a synthetic person, and internal production stages. Then test the labeled versions against unlabeled equivalents where the law lets you, and let your own numbers, not the panic, set the line.
Mikołaj Salecki
Writes about media, tech, and AI business for people who actually run digital. Former agency lead. Skeptic of frameworks that read better than they perform.
Get the next issue by email.
One letter, once a week. Sharp coverage of media, tech, and AI business. No filler.
