28 characters per AI-generated person.

New York's version of this costs $1,000. That's the number in every roundup you've read since Amazon's announcement last week.

It's the small one.

On Sunday, August 2, the EU AI Act's transparency rules take effect. Same imagery. Same synthetic models in your lifestyle shots and A+ modules. Different penalty structure: up to €15 million or 3% of your total worldwide annual turnover, whichever is higher. For small and mid-size businesses the ceiling is whichever is lower — which for an eight-figure seller lands around $240,000.

It applies to you even if you've never set foot in Europe. If your listings are live in DE, FR, IT, ES or NL, you're in scope.

Five days. Here's everything, in order.

The gap between the first two dates is 44 days. Every synthetic model you uploaded in that window is already sitting in your catalog untagged.

And to kill the rumor early: the EU's Digital Omnibus did delay parts of the AI Act. It delayed the high-risk system rules to December 2027. It did not touch Article 50. August 2 stands.

Amazon's help page — How to tag media that contains an AI-generated person, filed under Manage All Inventory → Create and manage offers — gives you the requirement in one line: add the contains-synthetic-performer keyword to the XMP dc:subject field before upload.

Read Amazon's own framing of why: "To comply with jurisdictions that require disclosure of AI-generated people in product media."

Jurisdictions. Plural. Amazon isn't building a house rule here. It's building one mechanism that answers to New York, to Brussels, and to whatever comes next — and handing you the operating cost.

Now the part that matters for Sunday: metadata alone is not legal disclosure.

The EU's guidance is explicit that disclosure has to be perceptible without technical aids — a human has to be able to see it. Machine-readable marking alone is not enough. New York's "conspicuous disclosure" standard points the same direction.

Amazon has signaled it will surface a shopper-facing AI indicator on affected listings, and that visible badge is what would carry the disclosure. But it is not on the help page and I have not seen it live on a PDP. So treat the sequence as two separate obligations, not one:

One thing runs both directions, though. An untagged synthetic image isn't a neutral omission. It's a positive assertion that the person in your image is real.

Read this first, because it's the part that silently fails. Amazon doesn't just want the string in the field. The help page specifies the metadata must contain an rdf:li of contains-synthetic-performer inside the XMP dc:subject rdf:Bag. Some tools write a flat string to dc:subject instead of a proper bag entry. That looks tagged in your file browser and reads as untagged to Amazon. Verify, don't assume.

Mac: Open in Preview → ⌘I (or Tools → Show Inspector) → Keywords tab → click + at the bottom of the pane → type contains-synthetic-performer → Enter.

Windows: Right-click → Properties → Details tab → in the Description section, the Tags field → type contains-synthetic-performer → OK.

Amazon calls this one out directly: the Subject and Tags fields in Windows write to different metadata locations. Use only the Tags field, and leave the Subject field unchanged. Putting it in Subject produces a file that looks tagged and isn't.

Bulk / command line — Amazon's own three-step sequence:

exiftool -G1 image.png | grep Subject

# (no output — Subject is empty)

exiftool -XMP-dc:Subject+=contains-synthetic-performer image.png

# 1 image files updated

exiftool -G1 image.png | grep Subject

# [XMP-dc] Subject: contains-synthetic-performer

On the += operator, straight from Amazon: it appends to the keywords bag without overwriting existing keywords. Running the same command twice writes the keyword twice. Duplicates don't break the tag, but they bloat the XMP packet.

The practical translation: use +=, never =. A plain = wipes everything already in that field.

Verify every file before upload. That third command is the check — run it, see the keyword echoed back from [XMP-dc] Subject, then upload. A failed metadata write and a skipped file look identical from Amazon's side, and identical from a regulator's.

1. Main images. Highest consequence. If Amazon pulls a non-compliant main image and you have no replacement queued, the listing can be suppressed from search entirely.

2. A+ lifestyle modules. The blind spot. Nearly every brand that built A+ this year put a synthetic model in a lifestyle module and hasn't thought about it since. A+ is explicitly in scope.

3. Video. Synthetic presenters count. So do synthetic people in b-roll.

4. EU marketplaces. Do these before Sunday. Everything else can slip a few days. This can't.

5. Everything uploaded since June 9. This is not a go-forward policy. Go backward.

Tag it:

Don't tag it:

The call when you're unsure: would an ordinary shopper assume a human stood in front of a camera? If yes, and none did — tag it.

The asymmetry decides it. Over-tagging costs you a badge. Under-tagging costs you $1,000, then $5,000, plus removal, plus suppression risk, plus a percentage of European revenue.

I'm going to give you the nuance here rather than the scary version, because you'll make a better decision with it.

Article 50(4) requires deployers to disclose "deepfakes." The Act defines a deepfake as AI-generated content that resembles existing persons, objects, places or events. A model who has never existed arguably isn't resembling an existing person.

That's a real gray area, and some counsel are reading it as a carve-out.

Three reasons not to build your week around it:

One. The Commission's May 2026 guidance says synthetic humans in advertising that are indistinguishable from real people likely require labeling. The guidance is non-binding, but it's the best signal available on how this gets read.

Two. New York has no such gray area. Its law covers digitally created media that appears as a real person — full stop. That one is live right now and it doesn't care whether your model exists.

Three. The carve-out costs more to litigate than the tag costs to apply. You are weighing a legal argument against 28 characters.

The generators are about to mark their own output. Article 50(2) requires providers of generative AI systems to mark outputs in a machine-readable format, detectable as AI-generated. New systems from August 2; systems already on market get until December 2. Read what that means for you: the tools you use to make these images are being legally required to embed a detectable marker in them. "Don't tag it and hope" has an expiration date, and it's on the calendar.

The badge is still unpriced. Amazon is expected to surface an AI indicator to shoppers on affected listings. Nobody has published conversion data on it — there's no benchmark, because it isn't visibly rolled out yet. On your top ASINs, replacing the asset with real photography may beat carrying an untested label. That's a decision worth making deliberately instead of inheriting. If you spot the badge live on a listing this week, send me a screenshot — I want to see how it renders before I make a call on it.

Liability sits on you, not Amazon. New York's law exempts platforms and publishers explicitly. Amazon built the mechanism and stepped off the hook. You're the advertiser. It's your name on it.

Twenty minutes across four folders, and you close a June 9 exposure, a July 23 platform requirement, and an August 2 regulatory deadline in one pass.

Do the EU listings first. Sunday isn't a soft date.

See you Friday, Dan

Have any questions? Grab a 15-minute slot here: book a time here.

Subscribe to The Elite Edge here.

— Dan