Amazon's New AI Label

Remember New York's AI disclosure law? Amazon has now done the predictable thing: product images containing photorealistic AI-generated people need to be tagged, with a consumer-facing indicator on listings "where applicable." Logical response. And the more I sit with the details, the more the interesting part isn't the policy. It's the enforcement — and specifically, how Amazon chooses to build it, because that choice tells you exactly whose problem this is. Spoiler: it's yours.

Two words. "Where applicable." Doing the heaviest lifting of any prepositional phrase in the history of marketplace policy.

Because here's what those two words mean: Amazon has to decide, listing by listing, across a catalogue of hundreds of millions of products — most sold by third-party sellers — which images contain a fake human and which contain a real one. There is no version of that decision that involves a person in a swivel chair squinting at photos of someone holding a spatula. It has to be done by a system. An AI system. Detecting AI.

To comply with a state law about artificial intelligence, Amazon has to deploy artificial intelligence to catch artificial intelligence. (If that made your brain do a little ouroboros loop, congratulations — you've found the problem.)

The Law That Drew a Line Nobody Can See

This stems from New York's disclosure law, in effect since June 2026, targeting "synthetic performers" — digitally fabricated likenesses that look like real people and stand in for human actors. The carve-outs are where it gets philosophically spicy.

The rule doesn't cover fictional or licensed characters. And — this is the bit that matters — it doesn't cover real people who've merely been altered by AI. So a genuine human model whose skin got smoothed, whose lighting got rebuilt, whose entire background got generated? Fine. A wholly synthetic person who never existed, conjured from a prompt? Not fine. Needs a label.

The line the law cares about isn't "was AI involved." It's "is the human in this image real." And that specific distinction is the one thing the entire generative image stack has been purpose-built to make impossible to answer.

The Watermark That Marks the Wrong Thing

Here's where the enforcement logic starts eating itself, on exactly the signal you'd expect Amazon to lean on.

The obvious way to classify at scale isn't pixel forensics — it's provenance. Read the metadata. Check for C2PA content credentials, the watermarking California already mandates for large AI providers. If an image announces its own origin, you don't need a detective.

Except a photograph of a real model retouched through a generative pipeline and a fully synthetic person conjured from a prompt can carry the exact same provenance flag. Both passed through an AI system. Both emerge stamped as AI-involved. The watermark records that a generative tool touched the file. It cannot record whether there was ever an actual person standing in front of a camera.

The signal that was supposed to make this clean can't separate the permitted case from the prohibited one — because the law drew its boundary inside the category the watermark describes, not around it. (It's a bit like using a smoke detector to work out whether someone's cooking dinner or committing arson. The detector knows there's smoke. It has no opinion on intent.)

The Classifier Problem

Strip the watermark away and what's left is worse. Now you're asking a classifier to look at a face and estimate — probabilistically — whether the human is real or generated. And it's a guess on the precise axis where generative models improve fastest: every increment of detection accuracy becomes training data for the next increment of evasion.

Amazon sits on both sides of this desk, because a share of the imagery being judged was made with tools Amazon itself provides to sellers and actively encourages them to use. The platform is being asked to referee the output of its own products, on a question those products are designed to make unanswerable.

So the classifier can't reliably do the thing the law needs done. Which raises the real question: knowing that, how does Amazon choose to build the enforcement anyway? Because it will build something. And the shape of what it builds is the whole story.

How Amazon Actually Plays This

The tempting conclusion is that a classifier this shaky, run by a platform scared of penalties, over-flags everything — your legitimate real-model photoshoot slapped with an "AI-generated people" tag while your fully synthetic competitor sails through. Tidy narrative. But it misreads whose problem this is.

Read the law again. The disclosure duty falls on the party who produces the advertisement. On a marketplace, that's the seller — you — not Amazon. The New York penalty, $1,000 rising to $5,000, isn't a bill Amazon is bracing to pay. It's a bill you are. That single fact changes what Amazon is optimising for. It isn't trying to minimise its own fine exposure, because it doesn't carry much. It's trying to make sure that when a regulator comes asking, the answer is "the seller attested, the seller tagged, the seller is responsible" — with a paper trail to prove it.

That points to a design, and it's worth speculating on because it's such a characteristically Amazon one: two tiers doing two different jobs.

The first tier is the seller-facing obligation, and here Amazon has every reason to be broad and aggressive. Mandating self-tagging, demanding attestations, running a classifier that kicks anything borderline back to the seller to confirm — none of this costs Amazon a cent of conversion, because none of it is visible to the shopper. What it produces is an audit trail and a liability shield. The classifier's real job, on this reading, isn't to identify fakes accurately. It's to generate a defensible record that Amazon acted, so exposure routes to the seller who tagged — or failed to tag — rather than to the platform that hosted them. Enforcement here is cheap for Amazon and lands entirely on you. Expect it to be expansive precisely because it's free to the party building it.

The second tier is the shopper-facing indicator — the "contains AI-generated people" label a customer actually sees — and here Amazon's incentive flips hard. This one costs conversion, on Amazon's own inventory, from which it takes a cut. So this is where "where applicable" earns its keep. Amazon can hold a broad, strict tagging obligation over sellers while surfacing the visible indicator narrowly, selectively, on its own conversion-protective terms. Strict where the cost lands on the seller. Quiet where the cost would land on Amazon.

Put the tiers together and you get the actual likely posture — not under-enforcement, and not indiscriminate over-flagging, but asymmetric enforcement. Amazon enforces hard in the direction that transfers risk, and stays soft in the direction that would cost it money, and those turn out to be two different surfaces. The seller carries the full legal weight of a broad, conservative tagging regime. The shopper sees a label only when Amazon judges it worth the friction. You get the obligation. Amazon keeps the discretion over visibility.

That's a worse spot for the seller than the doomsday version, not a better one. In the doomsday version the label at least gets applied evenly. Here it's applied strategically, by a party whose interests are not aligned with yours — maximal on your liability, minimal on Amazon's conversion.

If You're Actually Using Synthetic Models, This Is Your Section

Because the sellers this rule is genuinely for aren't the ones with real photoshoots. They're the ones already using AI-generated humans — the fabricated model holding the supplement bottle, the synthetic face in the lifestyle shot. If that's you, you're now standing in both tiers at once: you carry the legal exposure of the tagging obligation and you're the most likely candidate for the visible indicator if Amazon decides to show it. Two different risks, stacked, and you own both.

Start with the tag you control. Under-tag to dodge the consumer label and you've put yourself on the wrong side of the state's disclosure duty — the exact liability Amazon's design is built to route to you. Tag honestly and you volunteer for whatever the visible indicator does to your conversion. That's a genuine bind, and it's the one worth thinking hardest about, because it's not a compliance question with a clean answer. It's a bet on shopper behaviour.

And shopper behaviour almost certainly doesn't cut uniformly, which is the useful part. In commodity categories — cables, storage tubs, phone stands — a "contains AI-generated people" tag likely means close to nothing, because nobody was buying the human in the image anyway. In categories where the depicted body is part of the purchase logic — beauty, apparel, supplements, anything where a shopper is implicitly asking "will this work on someone like me" — a disclosure that the person isn't real plugs straight into the trust drawdown we've written about before. Not because one label is damning, but because it confirms a suspicion the shopper already half-held, at the exact moment of decision. Same tag, very different conversion consequences depending on what the human in the frame was doing for the sale.

None of this is settled, and anyone quoting you a firm conversion-hit number is guessing. But the shape of the decision is clear: synthetic humans are cheap to produce and now carry a stacked risk — a legal exposure you hold outright, plus a conversion exposure whose size depends on how much the realness of the person mattered to the buyer. In a category where it mattered, a real model, or no model at all, may already be the higher-converting choice before you even factor the label in.

What This Doesn't Mean

Before anyone spirals: most of what you do with generative AI sits nowhere near this rule. AI copy, fine. AI backgrounds, fine. Colour variants, product-on-white, lifestyle scenes with no fabricated human — all fine. The one creative choice now sitting under a forming legal spotlight is the photorealistic synthetic human likeness. That's the entire surface area.

So the move isn't to abandon generative tooling. It's to treat the synthetic person as the one input where the risk-reward has genuinely shifted — and, if you're leaning on it in a category where the human sells the product, to actually test the label's effect rather than assume you're safe or doomed. Run the synthetic version against a real-model version. Let your own conversion data settle a question the classifier can't.

The Direction Is the Point

We don't need to know exactly where this lands to read the trajectory. Amazon, holding a boundary it can't enforce precisely, a penalty regime pointed at its sellers rather than itself, and real conversion stakes on its own inventory, will resolve the whole thing in the direction that costs it least: aggressive liability transfer to sellers, discretionary visibility to shoppers. Not malice. Platform self-preservation math, which anyone who's sold on Amazon for five minutes will recognise as the governing logic behind roughly everything.

• Foundation argues branded Reddit persona accounts, human voices rather than logos, work as GEO assets. Reddit supplies roughly 21% of external citations across brands, so ranking comments feed both SERPs and LLM answers.

• ChatGPT's Work agent now signs into gated websites through a takeover cloud browser. Logins persist across sessions, so users authenticate once and let the agent keep working behind paywalls and portals.

• ChatGPT's agent market share falls from 78% to 56% in a year as Gemini doubles to 30% and Claude reaches 10%. Growth Memo names Google the most likely winner of the consumer AI market.

• TikTok tests TikTok Shop Plus, a Prime-style membership built into its Shop platform. US members get free shipping, product discounts, and coupons, borrowing Amazon's loyalty playbook to lock in repeat buyers.

• Perplexity launches Model Council in Computer, letting users convene 2 to 8 models from OpenAI, Gemini, Anthropic, and open source on one question. It synthesizes agreement and conflict, then ships reports and decks.

• Michaels' Gemini-powered assistant Ask Mike converts shoppers at more than double the rate of traditional search. Built on Google Cloud in six weeks, it drives 27% of chats to a product click or add to cart.

🖼️ Freepik Pikaso - A sketch-to-image AI image generator, in real-time.

🧠 Antispace - An AI tool that transforms thoughts into actions.

📹 Vidu 2.0 - A new video model with improved subject consistency and three resolution options up to 1080p

🎧 Syncly - Extract insights from customer feedback.

⁉️ My Ask AI - Let AI answer 75% of customer support emails.

😍It was great, all I needed

🙂Good, not great

🙄It sucked