Does AI-Written Product Copy Need a Label Under the AI Act?

A great deal of the product copy that European manufacturers write with an AI assistant this autumn will carry a mark they did not add and cannot see. Very little of it will need a label. The two statements sit together perfectly comfortably, and the widespread belief that every AI-written text must now be declared comes largely from running them into one.

The belief has a date attached, which lends it a certain authority. Article 50 of the EU AI Act, Regulation (EU) 2024/1689, sets out transparency obligations for certain AI systems, and it has applied since 2 August 2026. What it does not do, read line by line, is require a manufacturer to print "written with AI" at the foot of a product page for a hydraulic coupling. The claim is worth taking apart, because the parts of it that are true are not the parts most people are worried about.

First, the necessary caveat, stated once and plainly: this is not legal advice. It is a reading of the published text and of the Commission's own explanations, and a company with a real question about its own publications should put that question to counsel.

Two obligations, addressed to two different parties

Article 50 speaks to two roles, and most of the confusion comes from treating them as one. A provider is whoever develops an AI system and places it on the EU market: in this context, the companies that build the assistants. A deployer is whoever uses an AI system under its own authority, other than for purely personal, non-professional purposes. A manufacturer whose marketing team drafts a catalogue with Claude or ChatGPT is a deployer.

The duties differ in kind, not merely in degree. Providers must make the output detectable by machines. Deployers must, in certain narrow cases, tell people. The first is engineering done upstream, before the text ever reaches you. The second is a disclosure made by whoever publishes, and it is the second that a manufacturer actually needs to understand.

The provider's half, and what Claude now does with it

Article 50(2) requires providers of AI systems that generate synthetic audio, image, video or text to ensure the output is marked in a machine-readable format and detectable as artificially generated or manipulated. There is an exception where the system performs an assistive function for standard editing, or does not substantially alter the input the deployer supplied.

This is what sits behind the reports that Claude now watermarks its text, and on this point the reports are accurate. Anthropic's own documentation says that Claude models launched from 2 August 2026 embed an imperceptible watermark in the text itself, and that files of supported types carry signed provenance metadata. Anthropic applies this worldwide, across its API and its apps, on the stated grounds that it has no durable way yet to limit it by region. Models launched before that date are being brought in over a transition period.

For a manufacturer, the watermark changes rather less than the headlines implied. Detection is not a public checker that a customer or competitor can paste a catalogue into: Anthropic offers it in private preview to organisations eligible under EU law, such as regulators, media, fact-checkers and independent researchers. The mark carries no identifying information and cannot be traced to a person, an organisation or a conversation. It can indicate that Claude was likely involved with a text at some point; it cannot say who wrote which part, or whether some other system was involved. Light editing probably leaves it in place. Heavy rewriting, paraphrase or translation can remove it, which means a spec sheet drafted in English and translated into German may well lose the mark on the way, a detail exporters will receive with whichever emotion suits them.

Above all, the watermark is the provider discharging the provider's obligation. It does not, by itself, create a disclosure duty for the company that pastes the text into a product page. Anthropic's help centre is careful on exactly this point, telling organisations to assess independently what Article 50 requires of their own products and services.

The deployer's half, read slowly

Article 50(4) gives deployers two disclosure duties. The first concerns deep fakes: image, audio or video content generated or manipulated by AI, which has to be disclosed as such. Generated product imagery raises its own questions and is not the subject of this piece. The second concerns text, and it is the sentence on which the whole belief rests.

A deployer of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest must disclose that the text has been artificially generated or manipulated. The Commission's published answers on Article 50 break this into three conditions, all of which have to be met: the text is published, it is informative to the public, and it concerns a matter of public interest. The examples of public interest given there run from politics and democratic processes, public administration, justice and fundamental rights to public security, public health, environmental protection, consumer safety, and economic, financial, scientific or cultural developments that may be a relevant subject of public debate.

Now hold a product page up against that. A page setting out the flow rate, thread standard and operating pressure of a valve is published, certainly. It informs, after a fashion. But its purpose is to sell a valve to someone who needs one, and the list of public-interest matters reads like the contents page of a broadsheet rather than a parts catalogue. The Commission's answers say nothing about commercial or product content either way, so the careful position is that ordinary product copy appears to sit some distance from the provision. Not that it has been expressly excluded, because it has not.

Then comes the exception. Even text that does fall within the duty need not be labelled where it has undergone human review or editorial control and a natural or legal person holds editorial responsibility for its publication. The Commission describes human review as the deliberate examination of the substance of the content by people with relevant knowledge and professional judgement, and states that superficial or purely formal checks, spell-checking and grammatical correction among them, do not count.

Where the belief is not entirely wrong

Taking a claim apart is not the same as dismissing it. Three pieces of it survive the exercise.

Consumer safety is on the list. A manufacturer publishing a safety notice, a recall statement or general guidance on the hazards of a product class is closer to the provision than one publishing a datasheet. That is not an argument for labelling, but it is an argument against assuming that everything on a corporate website belongs to a single category.

Review has to be real. The exception rests on a person with relevant knowledge examining the substance. An engineer checking the tolerances an assistant has written into a specification is plainly that. Someone glancing for typos before pressing publish is plainly not, and a company that intends to rely on the exception would do well to be able to say who reviewed what.

The mark is there regardless. Whether or not a label is ever required, text drafted with an assistant that marks its output may carry a signal that eligible organisations can test for. No publishing process should be built on the assumption that AI involvement is undetectable. Article 50 is also, by its own terms, without prejudice to other transparency obligations in Union or national law, which is a reminder that it is not the only rule touching published content, and not the subject of this piece.

The dates that matter

Article 50 has applied since 2 August 2026, to providers and deployers alike. The Digital Omnibus on AI, in force since 27 July 2026, added a limited grace period: only for AI systems placed on the market before 2 August 2026, and only for the provider's marking and detection obligation under Article 50(2). Providers of those systems have until 2 December 2026. Nothing in that grace period touches a deployer's disclosure duty. Content generated before 2 August 2026 does not have to be labelled retroactively, although the Commission encourages it where possible.

The Commission has also published guidelines on the scope of Article 50 and a code of practice on marking and labelling AI-generated content, which is where the finer detail lives for anyone whose publications sit nearer the line than a datasheet does.

For completeness, since this is a vendor's blog: PSentry measures what AI assistants say about a brand in each language and market. It has nothing whatever to say about whether your content complies with anything, and should not be asked to.

What manufacturers ask once the alarm subsides

Do we have to label product pages we drafted with an assistant?

On the text of Article 50(4) and the Commission's three conditions, ordinary product pages do not obviously fall within the disclosure duty, and text that has had genuine human review benefits from the exception in any case. Whether a particular page is ordinary is a question for counsel rather than for a blog.

Can a customer tell that our catalogue was written with Claude?

Not with a public tool. Anthropic offers detection to organisations eligible under EU law, the result indicates only that Claude was likely involved, and heavy editing or translation can remove the mark altogether.

Can we switch the watermark off?

Anthropic's documentation does not describe any way for users to do so. It is the provider's compliance measure, applied across its products and in every region.

Does the grace period until December apply to us?

No. It covers providers of AI systems already on the market before 2 August 2026, and only their marking and detection obligation. A deployer's disclosure duty, where one exists, has applied since 2 August.

Would it be simpler to label everything, just in case?

That is a policy choice rather than something Article 50 asks for, and it has costs of its own. A label on every datasheet tells a buyer very little, and nothing at all about whether the figures are right. If a company does decide to label, it is worth being deliberate about where the label actually adds information.