I read Anthropic's actual contracts after the watermark announcement. Three things nobody is reporting: you can't sue as a class, liability is capped at 12 months of fees, and the Terms never mention marking at all.

I'm a paying commercial customer in Germany. I've spent several thousand dollars on Claude and I'm cancelling. Before I did, I went and read the things nobody reads. Posting what I found, because most of the coverage is missing the parts that actually matter.

1. Their own support page admits the mark can land on text YOU wrote.

> "Output can carry a Claude mark even if the underlying ideas, text, or data originated from another source."

That's from Anthropic's support doc . Read it again. Paste in something you wrote, ask Claude to proofread it, and the result can carry the mark.

EU AI Act Article 50 — the regulation they cite as the basis — is about AI-generated content. Your own writing is not AI-generated content. That isn't compliance, it's over-application, and it's the strongest objection anyone has here.

2. Their Terms never mention any of this.

Commercial Terms (effective 17 June 2025) and Consumer Terms (effective 8 October 2025). I searched both. Zero mentions of watermarking, marking, provenance, metadata or C2PA. What they do say:

> "Customer … owns its Outputs" — and Anthropic "hereby assigns to Customer its right, title and interest (if any) in and to Outputs."

So they assign you ownership of the artefact, then embed something into it, with no clause anywhere authorising that. Both contract versions predate the 11 August announcement.

3. The part that changes what you can actually do — and nobody is reporting it.

If you're in the EEA, the Commercial Terms put you under Irish law, with binding arbitration in Dublin (UNCITRAL rules). Not California. And:

- Class action is waived. "EACH PARTY AGREES THEY ARE WAIVING … THE RIGHT TO JOIN AND PARTICIPATE IN A CLASS ACTION." The "let's all sue together" plan does not exist contractually.

- Liability is capped at the fees you paid in the previous 12 months. Whatever you spent this year is your ceiling. Arbitration in Dublin costs more than that.

Individual legal action is economically dead on arrival, by design. Which is why the route below matters.

4. Switching to Chinese models does not get you out of labelling.

I keep seeing "just move to DeepSeek/Kimi". China's Measures for the Identification of AI-Generated Synthetic Content have been in force since September 2025, and Kimi and DeepSeek already document AI-content identification requirements under it. Different regime, same obligation, and it predates the European one. Move for price, self-hosting or latency if you want — but not because you think you're escaping marks.

What actually has traction

Complain to your national AI market-surveillance authority. It's free. In Germany the KI-MIG passed on 11 June 2026 and makes the Bundesnetzagentur the market surveillance authority and the central complaints body for the AI Act. A class-action waiver can't touch a regulatory complaint — it isn't litigation. Other member states have equivalents.

Critical: do not complain that they watermark. Marking is what the regulation tells them to do — that complaint gets binned. Complain that they mark content the user supplied, which is outside Article 50's scope and which their own documentation admits happens.

Send a written complaint before you cancel. An active account has a better support path than a closed one. Ask exactly these three, in writing:

- Is there any contractual opt-out or enterprise exception to output marking for commercial customers?

- Does text marking apply to source code via Claude Code and the API? Their doc lists .svg, .png, .jpg for file-level C2PA and says nothing about source files.

- On what legal basis is customer-supplied content marked, given Article 50(2) concerns AI-generated content?

Whatever they answer becomes the record. Whatever they don't answer is also the record.