Anthropic just made marking the default
On August 11, TechCrunch reported that Anthropic will watermark text generated by its AI models, Claude included. The stated reason is compliance with the EU's Transparency Code, which took effect on August 2. For every model released after that date, marking is automatic, and it covers both text and files. For files, Anthropic is using C2PA, the open provenance standard, rather than inventing a proprietary format.
That last detail matters more than the headline. A proprietary watermark is a vendor feature. An open standard is infrastructure: other systems can read it, verify it, and build policy on top of it.
The regulation behind the announcement
This is not a voluntary gesture. Article 50 of the EU AI Act requires that AI-generated or AI-edited content be marked in a way other systems can recognize, and the transparency obligations are in force as of August 2, 2026. Watermarking solutions themselves have until December 2, 2026 to reach compliance. So the window between now and December is when the tooling settles: model providers ship their marks, and downstream software learns to read them.
The scope is wider than many teams assume. Article 50 also covers the basics: disclosing that a chatbot is a chatbot, and labeling deepfakes. If your product talks to customers or publishes generated media, some part of this already applies to you.
What changes for content pipelines
If you run a marketing content pipeline, the practical question is simple: which tool stamps which mark, and where does that mark survive?
Most pipelines today are chains. A model drafts, an editor rewrites, a designer drops the image into a template, an export step compresses everything. Each hop can preserve provenance data or destroy it. Two consequences follow:
- Flows will split. Content that passed through meaningful human editing and fully automated content will diverge in how they are marked and how they should be labeled. Treating them as one bucket stops being viable.
- Your toolchain becomes a compliance surface. You need an inventory: which generators you use, what each one embeds since August 2, and what your CMS, image processor, and social schedulers do to that metadata on the way out.
Design provenance in, not on
The mistake to avoid is bolting labels on at the end. Provenance works when it is designed into the pipeline from the first step: record what was generated, by which model, and what a human changed, then let the label at the end be a byproduct of that record rather than a guess. When we build content automation at Clodron, that source chain is part of the architecture, not an afterthought, because retrofitting it across a live pipeline is far more expensive than starting with it.
What to do before December
- Inventory your generators. List every AI tool that produces text, images, or video in your pipeline, and check what marking each one applies since August 2.
- Test survival. Run a marked file through your real export path and verify whether C2PA data survives compression, cropping, and re-encoding.
- Separate your flows. Define which content classes are human-edited and which are fully automated, and decide labeling per class.
- Fix your supplier contracts. If agencies or freelancers deliver generated content, make marking and disclosure part of the agreement.
The providers are moving first. The December 2 deadline is for the tooling, but the operational readiness is yours to build, and the teams that map their pipelines now will spend the winter shipping instead of auditing.