A seller photographs her ceramics on the kitchen table because that is where the light is good. The photo is real, the pot is real, she made it with her own hands. The background is her kitchen, which is cluttered, so she uses the background replacement tool that came free with her phone's photo app and swaps it for a clean grey.
Under Etsy's current rules, that listing needs an AI disclosure. Without one it can be filtered out of search results, and she will most likely never be told why her views dropped.
That is not a hypothetical edge case designed to make a point. It is the single most common way honest sellers are falling foul of a policy they broadly agree with, and it is why roughly 68% of active Etsy shops had not updated a single listing within the first thirty days of the rule going live.
What actually changed
Etsy began enforcing AI disclosure on listings that use generative AI on 14 January 2026, adding a dedicated disclosure field inside the listing editor. The policy was tightened again through August 2026, with the added requirement that product images must reflect the actual item being sold rather than an idealised or generated version of it.
The enforcement mechanism is the part sellers underestimate. The disclosure field is not optional, and listings that do not complete it are filtered out of search results until the seller does. That is a quieter penalty than removal and in some ways a worse one, because a removed listing sends you a notification and a suppressed listing simply stops getting seen. A shop can lose most of its traffic to this without ever receiving a message explaining it.
The numbers from the first quarter give a sense of scale. Around 12,000 listings were removed and roughly 8,500 warnings were issued. Removal without prior warning is possible, which means the sequence some sellers expect, where you get told first and corrected second, is not guaranteed. For a shop where a handful of listings carry most of the revenue, that is a meaningful exposure sitting behind an unticked box.
What counts as AI involvement
This is where almost all the confusion lives, because Etsy's definition is deliberately broad and captures workflows most sellers do not think of as AI at all.
The obvious cases are obvious. An image generated in Midjourney, ChatGPT, or the old DALL-E needs disclosure. A product mockup produced from a text prompt needs disclosure. Patterns, SVGs, and digital downloads where the underlying design was AI-generated need disclosure. Nobody is surprised by any of that, and sellers working that way generally know where they stand.
The cases that catch people are the ones involving a real photograph of a real product. Background removal counts. Generative fill counts, which is what the ceramics seller used. AI-enhanced product photos count, which covers a wide range of one-tap improvements now built into ordinary phone and desktop photo tools. The policy captures any tool that swaps a background using generative fill even when the original shot was taken traditionally, which is the specific sentence most sellers have never read.
And it extends past images. Listing descriptions written by ChatGPT or any other language model require disclosure too. That one is worth sitting with, because writing product copy with AI assistance is now close to universal among sellers who have more listings than time, and a great many of them have never considered that the text field carries the same obligation as the photograph.
What you actually have to tick
The mechanics are simple enough once you know where to look, which is most of the problem, because the field was added to an editor most sellers navigate by muscle memory.
In the listing editor there is now a dedicated AI generation disclosure field, and completing it is the load-bearing step. Alongside that, listings involving AI generally need "Designed by" selected in the Item Details dropdown rather than "I made it", which is a meaningful distinction to Etsy and an emotionally difficult one for a lot of makers who did in fact make the thing. The distinction is about the design origin rather than the physical production, and reading it as an accusation is understandable but not what it means.
An explicit AI disclosure statement in the listing description is also expected. Plain language works better than legal language here: saying that the product image uses AI background replacement, or that the design was AI-assisted and hand-finished, communicates more to a buyer than a formal disclosure clause and reads as confidence rather than concession.
The practical approach for a shop with many listings is to work through them in order of revenue rather than in the order they appear. Your top ten listings almost certainly carry the majority of your income, and getting those compliant this week matters far more than a complete pass through two hundred listings you will abandon halfway. Compliance work follows the same rule as most business work: the part that pays is the part that touches the money.
How Etsy is finding out
The enforcement is not manual review, and understanding the mechanism explains why quietly hoping to be missed is a weaker strategy than it used to be.
Etsy matches what sellers declare against its own computer vision pipeline. In other words, the platform analyses the image independently and compares the result to the disclosure on the listing. A mismatch between what the system detects and what the seller declared is exactly the signal it is looking for, and it scales in a way human moderation never could.
The tools working against non-disclosure have also got considerably stronger this year, and not because of anything Etsy built. Images from ChatGPT now carry C2PA provenance metadata and an imperceptible pixel-level watermark on every tier, applied automatically. The metadata does not survive a screenshot, but the pixel watermark survives most ordinary editing. Anthropic shipped equivalent marking for Claude text in August, which we covered in the Claude watermark article. The generation tools are now labelling their own output, and the marketplaces are reading the labels.
The honest read on this is that non-disclosure is a strategy with a shrinking runway. It has probably worked fine for a lot of sellers so far. The mechanism that made it work, which is that detection was hard, is being removed by the AI companies themselves for reasons that have nothing to do with Etsy.
Why 68% of shops have not done it
The eRank figure, that roughly 68% of active Etsy shops had not updated a single listing within the first thirty days, is worth understanding rather than tutting at. The reason is not defiance.
Most sellers do not know the rule applies to them. If you take your own photographs of things you made with your own hands, you have every reason to assume an AI disclosure policy is aimed at somebody else. The gap between that reasonable assumption and a policy that counts background replacement as AI involvement is where the majority of non-compliance sits, and it is a communication failure more than a behaviour one.
There is also a genuine emotional obstacle in the "Designed by" versus "I made it" distinction. For a maker whose whole identity and pricing rests on handmade work, being asked to reclassify a listing feels like being asked to admit something untrue, and the natural response is to leave it and hope. That reaction is human and it is also the expensive option, because a suppressed listing does not argue back.
And the penalty is invisible, which removes the feedback loop that normally corrects this sort of thing. A seller whose listings are filtered from search sees a traffic decline with no explanation attached, and the available explanations are numerous: seasonality, competition, an algorithm change, bad luck. Compliance almost never makes that list, because nothing told them it should.
The wider lesson for anyone selling online
Etsy is the first major marketplace to enforce this seriously, and it is very unlikely to be the last. The direction across regulation and platform policy is the same, and it arrived from two directions at once this year.
Article 50 of the EU AI Act requires AI-generated content to be labelled with machine-readable marks, with synthetic image content due by 2 December 2026. The AI providers have begun marking their own output regardless of any marketplace asking them to. Those two facts together mean the infrastructure for detecting undisclosed AI content is being assembled by parties who are not marketplaces, and the marketplaces are simply reading what is already there.
The practical posture that survives all of this is to decide your disclosure position once and apply it everywhere, rather than tracking each platform's rules separately and updating reactively. A business that knows which of its images and text are AI-assisted, and says so plainly wherever it sells, does not need to monitor policy changes. One that depends on non-detection has to keep winning an argument that gets harder every quarter.
For the ceramics seller, the resolution is genuinely small. She ticks a box, adds a line saying the background was digitally replaced, and carries on selling pots she made with her hands to people who want them. The pots were never the problem. The unticked box was, and nobody told her it was there.