Claude output now carries an invisible watermark, and it marks processing, not authorship

Anthropic has signed the EU AI Act Code of Practice on transparency and switched on content marking for Claude. According to Anthropic's own documentation, Claude models launched on or after August 2, 2026 embed an imperceptible, machine-readable watermark in all generated text. Image files the model produces (.png, .jpg, .svg) carry digitally signed provenance metadata under the C2PA standard. The scope sentence is worth reading in full: marks apply "across Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag, and wherever Claude is offered, worldwide."

Three details in that documentation change what you should do this week, and none of them appear in the headline.

Whether your output is marked depends on the model date, not the product

The cutoff is the model's release date. Models launched on or after August 2 mark their output from day one. Models released before that date do not mark yet; Anthropic says marking support for them is "in progress" and gives no completion date.

That means a single product calling two Claude models through the same API key is currently producing a mix of marked and unmarked text, and the difference is invisible to you. If any part of your pipeline promises customers unmarked output, or conversely needs to demonstrate marking for its own Article 50 story, the first concrete step is an inventory: list every model id your product calls, place each one against the August 2 line, and note that the unmarked half of the list has an unannounced expiry date.

A proofread gets the same mark as a first draft

Anthropic is explicit about the central limitation: detecting the watermark proves a Claude model processed the text, not that Claude wrote it. Ask Claude to fix the grammar in a paragraph a human wrote, translate it, summarize it, or reformat it, and the output carries the mark. The mark answers "did a model touch this," never "who is the author."

This is the transferable part, and it outlasts anything specific to Claude. Every provider shipping text watermarking under Article 50(2) faces the same property, because marking attaches to model output and models spend much of their day processing human writing. Two policies to update accordingly:

• If your organization treats an AI-detection result as evidence that a person did not write something, that policy is now wrong by construction. A journalist who runs drafts through Claude for proofreading will produce marked articles they authored entirely.
• If you publish human writing that Claude has edited and your audience would read a detected mark as "machine-written," keep your own provenance: drafts, edit history, commit trail. The watermark cannot make the distinction for you, so your records have to.

Nobody outside Anthropic can read the text mark yet

There is no public detection tool. The documentation says detection mechanisms will be described "in forthcoming technical documentation." So today the text watermark is being written into output at scale while no customer, platform, or school can check for it. Marks accumulate now; detection arrives later, retroactively covering everything generated in between. If the presence of a future-detectable mark matters to your product, the time to decide is before the detector ships, not after.

Images are different. C2PA is an open standard, and the signed metadata on generated image files is inspectable today with standard tooling such as c2patool. Two practical notes there: metadata can be stripped by re-encoding or screenshotting, and if you are the one placing a generative feature on the market, stripping provenance from outputs is a decision to make with Article 50(2) in front of you, not a default in your image pipeline.

The text watermark has limits too. Anthropic's phrasing is that it survives copy and paste and "may persist through some editing." Heavy rewriting, translation by another system, or format conversion can degrade or remove it. It is a signal, not a seal.

December 2 makes this the default, not the exception

The regulatory context sits in one sentence: Article 50(2) of the EU AI Act requires providers of generative systems to mark output in a machine-readable way, systems already on the market before August 2, 2026 have until December 2, 2026 to comply, and Anthropic chose to apply its implementation globally rather than fence it to the EU. We walked through the full provider-versus-deployer split, the exemptions, and the fine structure when the transparency obligations took effect on August 2.

The practical consequence of the deadline: OpenAI, Google, and every other provider serving EU users need an equivalent answer within roughly sixteen weeks, and there is little engineering reason to build it EU-only. Marked output is on its way to being the default state of model-generated text. The inventory question, the processing-versus-authorship correction, and the decision about your own provenance records will apply to each provider as they ship, so the work you do on the Claude case this week is reusable.

Model behavior changing underneath a stable API is not a new pattern; we covered the version of it where model ids change under you. This is the quieter version: same id, same latency, same token bill, different output properties. It is worth a line in your vendor-watch notes either way. If you build on Claude or run Claude Code in a pipeline that emits text to customers, the output you shipped yesterday and the output you ship today are no longer the same artifact.