Anthropic Puts Invisible Watermarks on Claude's Output — a Blanket Mark Driven by the EU AI Act

Anthropic has begun embedding invisible watermarks in text and images generated by Claude. The move complies with the EU AI Act transparency obligations that took effect on August 2, applies automatically to models released after that date, and covers every Claude product — API, Claude Code, Cowork and more — with no opt-out. Text watermarks survive copy-paste and some editing.

Anthropic is embedding invisible watermarks in text and files generated by Claude. It's a step toward compliance with the EU AI Act's transparency code, which took effect on August 2. The watermark applies automatically to models released after that date and will expand to older models later. Text watermarks travel with copy-paste and survive some editing; images are marked with C2PA standard metadata. With no opt-out, some users are pushing back, worried their work will be flagged as AI-generated.

Anthropic has confirmed that it is starting to embed watermarks — invisible to the human eye — in content generated by Claude. On August 11, several outlets reported the change, citing a support article Anthropic had updated the day before. The driver is the EU AI Act's transparency obligation, which took effect on August 2 and requires AI-generated or edited content to be marked in a way other systems can identify.

What, and from when

According to Anthropic, Claude models released after August 2 will automatically watermark generated text and files. For files, the company uses the open C2PA (Coalition for Content Provenance and Authenticity) standard. Anthropic says it will extend the capability to older models over time. Crucially, watermarking is applied at the model level — meaning the mark travels no matter which Claude product or surface the text comes from.

Effective date Aug 2, 2026 (EU AI Act transparency code)
Covered models Automatic for models released after Aug 2; older models to follow
Image standard C2PA (JPG, PNG, SVG and more)
Opt-out None — applied across all products, worldwide

How the text watermark persists

The support page explains that "because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing." Anthropic did not spell out exactly how it works, but the industry expects an approach similar to the watermarking method Google DeepMind published in Nature in 2024. That technique subtly biases a language model's word choices so the output becomes statistically detectable over enough content — DeepMind's SynthID being the best-known example.

Scope and the limits of detection

The watermark applies across products including the Claude Platform API, Claude, Claude Code, Claude Cowork and Claude Tag. For images, C2PA attaches metadata specifying which model created a file and when, plus any copyright restrictions, and uses a hash to detect tampering.

Target	Marking method	Notes
Text	Word-choice-bias watermark	Survives copy-paste and some editing
Images / files	C2PA metadata	Records model and time; hash for tamper detection
Detection tools	Coming from Anthropic	May miss short or heavily edited text

Anthropic says it will release tooling to help detect Claude-generated content, while cautioning that it "won't be perfect." In particular, it may not always catch watermarks in short or heavily edited snippets of text.

C2PA pairs an AI image with a metadata file recording who, when and which model created it, along with copyright information. It generates a unique hash for each metadata file to spot edits easily, and is designed to detect even attempts to swap out the entire metadata file rather than merely edit it.

Backlash and the industry trend

Because there is no opt-out, some users are objecting that the mark will reveal they used Claude at work or in class. The main complaints: their own output being categorically stamped as "AI-generated," the risk of false accusations, and a lack of technical implementation detail.

If you draft with Claude, the habit of substantially rewriting and editing the final output matters more than ever. But editing purely to strip the watermark raises its own integrity questions — it's safer to treat AI use as a given and lean into review and rewriting, rather than trying to erase the mark.

This is not Anthropic acting alone. AI music platform Suno recently said it would mark its tracks, and Black Forest Labs, Google, Meta, Microsoft, OpenAI and Synthesia have all committed to adhering to the EU code. OpenAI sits on the steering committee of the consortium that develops C2PA, so it may launch its own text watermarking soon.

What it means

The move signals that regulatory and industry responses to "AI slop" and the broader trust crisis are now showing up in shipping products. With the EU pinning a hard date (August 2), major model providers are implementing marking obligations in lockstep. The open question is detection accuracy versus evasion: the fact that the watermark is weak against short text and heavy edits lays bare the gap between the regulatory goal (identifying provenance) and the reality (easy circumvention).

Anthropic is embedding invisible watermarks in Claude-generated text and images to comply with EU AI Act transparency obligations.
Automatic for models released after August 2, expanding to older models later. No opt-out; applies across API, Code, Cowork and other products.
Text survives copy-paste and some editing; images are marked with C2PA metadata — but short or heavily edited text may evade detection.
The lack of an opt-out has drawn user backlash. Google, Meta, Microsoft and OpenAI have made the same EU commitment, pointing to industry standardization.