Anthropic Watermarks Claude Text and Images for EU Rules

Anthropic will now embed invisible, machine-readable watermarks into every piece of text and image Claude generates — a compliance move some users are calling a "digital tattoo."

The announcement, posted on a newly created Claude support page, describes two mechanisms working together: invisible watermarks woven into the token stream of generated text, undetectable to human readers but extractable by specialized tools, and digitally signed provenance metadata attached to generated files in formats that support it, such as PDF, PNG, or JSON-L.

How the Watermark Works

For text, the model selects a secret key, derives a pseudo-random sequence from it, and nudges token sampling toward choices that encode a binary pattern — adjustments that are, in Anthropic's framing, "often a fraction of a probability point," so the output reads like any other Claude-generated passage. A detection tool later reverses the process, checking whether the token distribution matches the expected pattern. For images, Anthropic plans to embed a spread-spectrum signal into the pixel domain or the diffusion model's latent representation, designed to survive compression and resizing.

Signed Provenance Metadata

The provenance side works differently: when Claude finishes a request, it generates a JSON-L block with fields including model_id, generation_timestamp, prompt_hash, and output_hash, then signs that block with Anthropic's private key using an algorithm like Ed25519. Anyone holding Anthropic's public key can verify the metadata hasn't been altered. For plain-text files that can't hold embedded metadata, Anthropic plans to ship the signed block as a side-car file instead.

Why Now: EU Compliance

Anthropic is calling the rollout a "future commitment" — the technology isn't active yet. The trigger is the EU AI Act's Transparency Code, which requires tech companies to label AI-generated or AI-edited content in a way identifiable to computer systems; noncompliance can carry fines of up to 6% of global turnover.

The Backlash

Not everyone is on board. On Reddit, one user argued the watermark will catch ordinary users, not misuse: "Who will get caught? You. The student who used Claude to reorganize a paragraph. The journalist who asked the AI to summarize a two-hundred-page transcript." Another objected to being credited alongside the tool: "I gave the instructions, context, decisions, and countless refinements, claude was the tool. If Claude starts watermarking the code or anything else it generates, what exactly is it claiming credit for?" Other commenters pushed back just as hard, with one calling the criticism overblown: "It's not claiming credit though. It's about being able to detect AI generated outputs because of the risks AI generated outputs can cause in various situations."

What to Watch

Anthropic isn't alone here — OpenAI has tested textual steganography and Stability AI ships a diffusion watermark for images, though Anthropic's combination of invisible text watermarking with signed provenance metadata is more layered than either. Since the system isn't live yet, the real test comes when detection tools ship: whether they stay accurate after paraphrasing, and whether other major labs follow with matching disclosure standards under the same EU deadline.