INITIALIZING…
SYSTEMS
00
00
Tushaar Naagar
Back to writing

AI

Claude Is Watermarking Output. Here's What Changes If You Ship That Code

Aug 12, 20262 min read

Anthropic is rolling out statistical text watermarks and C2PA image manifests worldwide to meet EU AI Act transparency rules. Detection is about to get boring — and official.

The policy, not the panic

Anthropic signed the EU Code of Practice on Transparency of AI-Generated Content. Article 50 of the AI Act started applying on August 2, 2026: generative output has to be marked in a machine-readable way. Claude now embeds a statistical watermark on text and C2PA-signed provenance on supported images (PNG, JPG, SVG), across API, apps, Claude Code, and related products — worldwide, because they couldn't scope it to the EU cleanly. Models launched on or after August 2 ship with it on. Older models are on a transition window. Detection tooling is coming so platforms can check Claude-generated text instead of playing 'GPTZero roulette.'

What it is not

It is not a virus in your repo. Statistical watermarks on text are meant to be robust enough to detect, not to brick a file. C2PA on PNG/JPG/SVG is a signed manifest in the container: inspectors can see that Claude generated or processed the asset. Other labs — Google, Microsoft, OpenAI — are on similar Code of Practice tracks. Provenance is becoming a layer, like HTTPS did: invisible until you look, then expected.

What I'd do on a team that uses Claude Code

Don't treat watermarking as a reason to hide AI use. Treat it as a reason to write down where generated code is allowed: scaffolding and tests yes, auth and payments only with review. If a customer or a university starts checking provenance, you want a policy, not a surprise. For product UIs that emit model text, assume detection will exist and be cheap. Label generated copy when it faces a user. Keep human review on anything that hits production data.

Takeaway

Watermarking is a compliance feature that will leak into engineering process. The teams that look adult are the ones who already review AI-authored diffs like junior PRs — because that's what they are, with or without a cryptographic receipt.