Do you have to disclose AI product descriptions?

No, you almost certainly do not have to disclose AI product descriptions. The EU rule that people are worried about applies to text “published with the purpose of informing the public on matters of public interest”, and a listing for a linen shirt is not that. It also exempts anything that went through human review with someone holding editorial responsibility, which is what a merchant editing a draft actually does.

The question came back into circulation this week because Anthropic announced on 14 August 2026 that Claude now watermarks the text it produces. Cue a wave of merchants wondering whether their product copy is about to be flagged, penalised or exposed.

It is not. But the announcement is a useful excuse to sort out which AI obligation is whose, because there is one that genuinely lands on you, and it is not the one everyone is asking about.

In this post

• What Anthropic actually announced
• Why every major model will end up doing this
• Does any of it put a duty on you?
• Will Google punish AI written copy?
• Where you do owe something: images
• What to actually do this week
• FAQ
• Related reading

What Anthropic actually announced

Short version, from Anthropic’s own explainer: future Claude models produce text carrying a watermark, and the reason given is compliance with the EU AI Act. The method is described as a version of the SynthID-Text approach published by Google DeepMind, and it works by modifying “the source of the randomness used to pick among words” rather than changing which words get picked.

The properties that matter to anyone publishing that text:

• Nothing is added to the text and there are no hidden characters. No zero width spaces, no invisible unicode, nothing to strip.
• No extra tokens, no extra cost. The watermark does not change your bill.
• No quality impact. Anthropic reports no effect on content, creativity or readability, and says the difference will not be distinguishable to readers.
• Not traceable to you. It identifies the model’s output, not a person, an organisation or a conversation.
• Detection needs the key. Only parties holding the key that encodes it can detect the mark, and a detection API is described as coming.

That last point is the one worth sitting with, because it kills most of the anxiety in one line. A watermark nobody can read without a key is not a public badge on your product page. It is closer to a batch number on the inside of a seam.

Why every major model will end up doing this

This is not a company deciding to be careful. It is Article 50(2) of the EU AI Act, which requires providers of AI systems generating synthetic content to ensure the outputs “are marked in a machine-readable format and detectable as artificially generated or manipulated”.

Anthropic says it signed the EU Code of Practice on Transparency of AI-Generated Content in July 2026. That timing fits: the first signing wave closed at the end of July, and the Commission’s assessment of that Code landed in early July. Roughly 190 organisations had signed by then, and the provider section reads like a roll call of the model labs. So yes, the others are on the same hook.

Anthropic also notes a transition period for models launched before 2 August 2026. That is Article 111(4), added by the Digital Omnibus, giving providers whose generative systems were already on the market until 2 December 2026 to comply with the marking duty. Which means the rest of the industry’s announcements are due in the next few months, not next year.

Does any of it put a duty on you?

Article 50(2) binds providers. You are a deployer: you use somebody else’s model. The deployer duties sit in Article 50(4), and there are two of them, one for deepfakes and one for text.

The text one, verbatim: “Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated.”

Then the exemption, also verbatim: the obligation does not apply “where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content”.

A product description fails the trigger twice over. It is commercial copy, not public interest reporting, and if you read the draft before it went live, a human held editorial responsibility for it. Two independent reasons you are outside the rule.

What you published with AI	EU disclosure duty on you?
Product descriptions, collection copy, size guides	No
Marketing emails and ad copy	No under Article 50, though advertising law still applies
A blog post about your niche	No, commercial content with your editorial responsibility
A news style article on a matter of public interest, published unreviewed	Yes, this is the case the rule was written for
A chatbot on your storefront	Separate rule: people must be told they are talking to an AI
An AI product image that misrepresents the real product	Yes, and this is the one merchants keep missing

Note the shape of that table. Everything in the text column resolves to no. The only rows with a yes are a chatbot and an image.

Will Google punish AI written copy?

The watermark is not a Google signal. It is key-gated by design, so a search engine cannot read it any more than a shopper can, and it carries no identity to attach a penalty to. Whatever you think about AI written product copy, this announcement does not change your exposure in search by one point.

What does matter is whether the copy is worth reading, which is the same question it was two years ago. Google published its guidance about AI-generated content in February 2023 and it is short enough to read for yourself rather than through three layers of blog paraphrase, which is what I would recommend before you let anyone tell you what Google thinks.

My own position, for whatever it is worth: the risk with AI product copy has never been detection. It is that four hundred descriptions generated from the same prompt read like four hundred descriptions generated from the same prompt, and that shows up in engagement long before it shows up in a policy. Run your store through our free AI readiness checker if you want to see how machine readable your content actually is, which is a more useful question than whether it was machine written.

Where you do owe something: images

Here is the twist, and it is the reason this post exists rather than a one line reassurance.

The other half of Article 50(4) covers deepfakes, defined in Article 3(60) as AI generated content resembling “existing persons, objects, places, entities or events” that “would falsely appear to a person to be authentic or truthful”. The Commission’s guidelines of 20 July 2026 read “objects” as including consumer goods, and give the worked example directly: an AI generated image of a product that misleads “as to the actual product appearance, characteristics or use” is a deepfake. A real product against an AI background is not.

And the provider’s marking does not cover you. The same guidelines say deployers “cannot rely on the machine-readable marking embedded in the content by the provider under Article 50(2)”, because that marking is not clear and distinguishable to the people looking at the image. So the watermark your image tool embeds is its compliance, not yours. Yours is a visible label, and the penalty band for getting it wrong is up to 15 million euro or 3 percent of worldwide turnover under Article 99(4).

Doing that by hand across a catalog is where it stalls, which is why bulk tooling matters here: Viking Watermark applies an AI Generated label to selected photos by collection, tag or product status and keeps the originals in Shopify Files for a one click restore. We checked every app in both relevant App Store categories, 442 listings, and it is the only one shipping a ready made AI label. Free for the first 100 images, $5 a month above that, 5.0 stars from 6 reviews, which is a new listing rather than a proven one.

The full breakdown of which images are caught, what the label should look like and where it goes is in the EU AI Act guide for product photos.

What to actually do this week

1. Keep reading the drafts. Human review plus editorial responsibility is your exemption on the text side, and it is also the difference between copy that sells and copy that fills a box. You were going to do it anyway. Now it is load bearing.
2. Do nothing about the watermark. There is nothing to strip, nothing to disclose, nothing to budget for.
3. List which product images were generated rather than photographed. That list is your actual compliance scope, and it is usually shorter than merchants fear.
4. Do not strip image metadata. Google Merchant Center requires AI generated feed images to carry metadata indicating they were AI generated, and separately tells you not to remove it.
5. Keep the feed image clean. Ordinary overlays cause product disapprovals. Google’s own AI content labels are the documented exception. Check what your feed is sending with the free Google Shopping feed checker.

And while you are in the catalog: AI copy is not the thing costing you conversions on a product page. Showing the wrong photo when a shopper picks a colour is, which is variant image filtering work, and so is a collection page listing the same jacket in six near identical cards, which is what combined listings exist to fix.

FAQ

Do I have to label product descriptions written with AI?

Not under the EU AI Act. The text duty applies to content published to inform the public on matters of public interest, and it is disapplied where a human reviewed the content and holds editorial responsibility. Ordinary product copy misses on both counts.

Does the Claude watermark make my copy detectable as AI?

Only to someone holding the key that encodes it, and Anthropic says a detection API is coming. It identifies the model’s output rather than any person, organisation or conversation, so it cannot be traced back to your store.

Are there hidden characters I should strip before publishing?

No. Anthropic states that nothing is added to the text and there are no hidden characters. The method changes the source of randomness used when selecting words, so there is nothing in the output to find or remove.

Will watermarking make the API more expensive?

No. It does not require extra tokens and Anthropic says it will not be more expensive. Output quality is unaffected too, with no reported impact on content, creativity or readability.

Do the same rules cover my AI generated product photos?

No, images are the other half of Article 50(4) and they are stricter. An AI image that misleads about the real product’s appearance, characteristics or use needs a visible disclosure, and the Commission’s guidelines say the tool’s machine readable marking does not discharge that duty for you.

Does my storefront chatbot need a disclosure?

That is a different paragraph of the same article, and yes: people interacting with an AI system have to be told, unless it is obvious to a reasonably observant person. Most chat widgets already say it in the greeting.

Related reading

• EU AI Act: do AI product photos need a label?
• How to add an AI Generated label to product images
• The 2026 AI image disclosure compliance calendar
• Shopify variant images FAQ
• Shopify combined listings explained

So: nothing to do about your product copy, and one afternoon of work on whichever product images were generated rather than photographed. That is the whole compliance story, and it is a lot smaller than the week’s headlines suggested.