Written by: Nicole Ogonowska, IT Growth Manager, Digital Colliers
Last week Amazon started notifying third-party sellers that product ads using AI-generated people need a disclosure label to comply with New York's synthetic performer rules. That notice is doing something quiet but important. It's pushing a compliance obligation down from the platform to the seller, and from the seller to whoever manages creative assets. Which, on most eCommerce teams, is a Dropbox folder and a person named Sam.
The uncomfortable part: your DAM almost certainly cannot filter assets by "contains synthetic performer." That field doesn't exist. Nobody's been tagging for it. And the deadlines to fix that are not theoretical anymore.
What a compliant asset record actually looks like
If a regulator, a platform, or your own legal team asks tomorrow which live ads contain AI-generated people, you need to answer in minutes, not weeks. That means every creative asset needs structured metadata, not a filename convention.
The minimum record I'd argue for:
- Source model: which generator produced the image or video (Midjourney, Sora, Runway, an internal fine-tune, etc.), plus version.
- Synthetic performer flag: boolean, but with a sub-type. Fully synthetic person, AI-modified real person, AI-generated voice, or none.
- Human-in-the-loop record: who reviewed it, when, and against which checklist.
- Rights and consent trail: if a real person's likeness was used as training or reference input, the paperwork lives with the asset.
- Ad ID mapping: every downstream Meta, TikTok, Amazon, and Google ad ID that used this asset, so a takedown or relabel is one query.
- Disclosure text applied: the exact label copy, per jurisdiction.
Notice that last one. New York's rule is not the same as what the EU is about to want, which is not the same as what California will probably pass next. The label copy is jurisdiction-scoped, so the asset record has to hold a small matrix, not a single string.
Why spreadsheets break at 40 SKUs
A single-market brand with a dozen hero SKUs can just about survive with a tracker tab. The moment you're running 40 plus SKUs across UK, EU, and US, with seasonal refreshes and localized variants, the math falls apart.
Each SKU spawns 5 to 15 creative variants per market. Multiply by 3 markets. Multiply by refresh cadence. You are now looking at thousands of assets, each of which may or may not contain a synthetic performer, each of which is stitched into some ad set somewhere.
This matters more now because the pressure to ship AI creative is real. DTC customer acquisition cost has risen roughly 40% since 2023, and Meta CPMs kept climbing through 2024 and 2025. Teams are leaning on generative tooling to keep creative volume up without burning studio budget. That's rational. What's not rational is doing it without the metadata layer to survive an audit.
Spreadsheets also fail the basic test of "pull every live ad that uses model X, because model X just got sued." You cannot join a spreadsheet to your ad platform in real time. Your DAM can, if you set it up to.
The hard deadline nobody's putting on the roadmap
New York is the opening move. The EU AI Act's Article 50 transparency obligations, which cover disclosure of AI-generated content interacting with people, apply from 2 August 2026. That's the one to circle. High-risk obligations follow on 2 December 2027. Fines for the serious tier reach up to 15 million euros or 3 percent of global turnover, and GDPR-style enforcement culture means the fines are not decorative. GDPR itself already tops out at 20 million euros or 4 percent.
If you sell into the EU, August 2026 is roughly the horizon where "we'll sort it later" stops being viable. And if a second US state copies New York in the next 12 months, which is the direction of travel, the same asset needs two labels, tracked separately, per market.
What the operators shipping this in 2026 are doing
The pattern I keep seeing on teams that will not be panicking next year:
- They picked one DAM and forced every creative pipeline through it, including agency deliverables. No shadow folders.
- They added the synthetic performer schema before they needed it, and backfilled the last 18 months of assets.
- They wired the DAM to their ad platforms so "which live ads use this asset" is a query, not a project.
- They wrote the human-in-the-loop step into the workflow, so the reviewer signoff is a field, not an email.
None of that is glamorous. It's data plumbing. But the brands that treat AI creative as a data problem, not a design problem, are the ones who'll still be running ads on Amazon when the next state passes its version of the New York rule.