Your AI Clause Just Became Testable

Most of the conversation about AI watermarking is about students and journalists. I want to talk about what happens when a client runs a detector against a deliverable, because producers and agencies are going to deal with that first.

Anthropic published a page describing how it will mark AI-generated content. If you sign statements of work with AI clauses in them, and at this point a lot of us do, this is worth twenty minutes of your attention.

What was actually announced

Anthropic signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content. Their implementation, in their own words:

- Claude models launched on or after **August 2, 2026** support marking at launch. Models released before that date are still in progress. - Generated text carries an imperceptible embedded watermark. It survives copy and paste and "may persist through some editing." - Generated files like .svg, .png and .jpg get signed provenance metadata using the C2PA standard. - Marking happens at the **model level**, so it applies across the API, Claude, Claude Code, Claude Cowork and Claude Tag, and through AWS, Google Cloud and Microsoft Foundry. - It applies **WORLDWIDE**, wherever Claude is offered. - Detection tooling is promised but has not shipped yet.

On timing, the regulation gives you slightly more room than the headlines suggest. Article 50 applies from August 2, 2026, but systems already on the market before that date have until **December 2, 2026** for the marking obligation. Content generated before August 2 never has to be labeled retroactively. Penalties top out at 15 million euros or 3 percent of worldwide turnover.

The sentence that should worry you

Buried in the limitations section of Anthropic's own page:

> People often use Claude to proofread, translate, summarize, or convert files. The output can carry a Claude mark even if the underlying ideas, text, or data originated from another source.

So a document you wrote, that you ran through Claude to fix your commas, can carry the same mark as a document Claude wrote from nothing.

The reverse is where people get confused, so let me be plain about it. If a detector finds no mark, that tells you nothing. It does not mean a human wrote it. Anthropic lists the ways detection fails: the model predates marking, the text was paraphrased or translated or mixed into other writing, the passage was too short, or a file's metadata got stripped by a format conversion or a screenshot.

So the tool can flag you, and it can never clear you. Keep that in mind for everything below.

What the law exempts and Claude marks anyway

Here is what makes this a contracts problem instead of a philosophy problem.

**Copyright law already distinguishes assistance from authorship.** The US Copyright Office has been clear that work generated entirely by AI is not copyrightable, that only human contributions are protectable, and, critically, that using AI as a tool to assist a human creative process does not make the resulting work uncopyrightable. It is a case-by-case analysis of what the human actually did.

**The regulation draws a version of the same line.** Article 50(2) states the marking obligation "shall not apply to the extent the AI systems perform an assistive function for standard editing or do not substantially alter the input data provided by the deployer or the semantics thereof."

The mark draws neither line. It is applied at the model level to all generated text.

Now, the fair reading of why. That exemption is scoped **per use**, not per system. Claude does generation and assistive editing inside the same product, sometimes inside the same conversation. A model cannot reliably know whether the request in front of it is "rewrite this pitch" or "fix my commas," and a marking scheme that fired only on the first case would be trivial to defeat by phrasing your prompt differently. Marking everything is probably the only implementation that actually works.

So this is not a case of a company overreaching. The law asks what you used Claude for. Anthropic can only control what the model does on every single request. Nobody has figured out how to connect those two, and Anthropic is also doing more than the law strictly requires, which is not a violation of anything.

Either way, your client ends up with a tool that tells them less than their own contract needs to know.

I should also flag what I do not know. The Commission says its Guidelines on Transparency of AI-Generated Content contain "a diverse set of practical examples of what can be considered standard editing and what goes beyond." I have not read those examples, and the boundary is going to be fact-specific. Anyone telling you confidently where the line falls between a grammar pass and a substantial alteration is guessing.

What the regulator actually left out of scope

This part deserves more attention than it has gotten, because the carve-outs are broader than most of the coverage suggests. The Commission FAQ lists outputs that fall outside the marking obligation, including:

1. Short sequences of numbers, symbols or letters. 2. **Source code.** 3. Outputs communicated only machine to machine and processed automatically with no human exposure. 4. Outputs used only in closed-loop industrial and product development environments, **"for example for film production," unless they are the final output.**

Read number two twice if you ship code. Read number four twice if you make anything. The regulator explicitly contemplated that a production pipeline generates enormous volumes of intermediate material that nobody outside the process ever sees, and it put that material outside the obligation. Previs, temp comps, concept frames, layout passes, scratch tracks. Final output is a different question. The Commission put the working material outside the rule.

The mark does not make that distinction. Your previs frame and your delivered master come back the same.

The specific way this bites a producer

Think about a normal SOW clause. Some version of these is now common:

1. Vendor shall not use generative AI in the creation of Deliverables. 2. Vendor shall disclose any use of generative AI prior to use. 3. Vendor warrants that Deliverables are original works.

There have been AI detection programs out there but they could struggle to be 100% right. Once Anthropic ships its detection tooling, for supported models, a client can check.

And that detector cannot tell the difference between a document Claude wrote and a document Claude proofread. Both come back the same.

Files work differently from text. C2PA metadata rides on the file itself, so design assets carry provenance in a way a paragraph does not. Metadata is also the easiest thing to lose in a normal production workflow. Re-save a file, convert a format, take a screenshot, and the signal is gone. So whether you look compliant can come down to whether somebody exported a PNG twice.

What I would actually do about it

I am a producer, not a lawyer, and none of this is legal advice. But legal teams are going to have to sort this out, and sooner rather than later.

1. **Inventory your active SOWs for AI language.** You cannot manage a clause you have not read since signing. Sort them into prohibited, disclosure-required, and approval-required.

2. **Insist that "AI-generated" and "AI-assisted" are defined separately.** A contract that just says "no AI" is now a contract you can fail by using a spell-checker. Definitions are the whole ballgame here.

3. **Ask for language stating that a machine-readable mark is not by itself evidence of breach.** This is the rare case where the vendor's own documentation is your leverage. Anthropic says plainly that a detected mark is "not fully conclusive" and indicates only that content "may have been processed." Quote them.

4. **Ask whether intermediate work product is even in scope.** If the regulator carved out closed-loop production environments and source code, your contract can draw the same distinction between working files and final deliverables. Most SOW language I have seen does not.

5. **Push disclosure obligations down to your subcontractors.** You are warranting work you did not personally produce. If a freelancer used an assistant on a caption file, that is now your representation, not theirs.

6. **Keep a generation log per deliverable.** If you ever have to establish that a mark reflects assistance rather than generation, you will need contemporaneous records of what got made how. Reconstructing it after the accusation is not a position you want.

7. **Do not sign a warranty you cannot verify.** That was always good practice, and it just got a lot easier to get caught.

The bigger thing underneath

There are at least six different things a person can do to one document with Claude. Generate it from nothing. Rewrite most of it. Lightly edit it. Translate it. Convert the file. Review and approve it after the fact. Those six carry six different answers for authorship, for copyright, and for whether you breached your agreement. The mark reports all six exactly the same way.

Provenance is a good idea and I want a world with better chain of custody. A useful system would tell you which of those six things happened. This one only tells you that Claude was involved.

Paying attention to AI clauses and aligning with them is always crucial. Today's news takes that work up an extra level.