Claude Now Watermarks Every Text Response and File It Generates

As of August 2026, Claude marks the text and files it generates, worldwide, with no way to turn it off. If you or anyone on your team uses Claude to draft ad copy, landing page content, blog posts, or client-facing documents, it's worth understanding exactly what "marked" means here, because it's less dramatic than it sounds and also not nothing.

Why this is happening now

This isn't a feature Anthropic dreamed up unprompted. It fulfills a commitment under the EU AI Act's Article 50 Code of Practice on Transparency of AI-Generated Content, which Anthropic signed in July 2026 alongside roughly 190 other signatories. The Code requires AI providers to mark generated content so it can be identified as such. Anthropic rolled the marking out globally rather than just in the EU, which is why it applies to Claude accounts everywhere, not only European ones.

What actually gets marked, and how

Two separate mechanisms, for two different kinds of output:

• Text gets an invisible watermark. Anthropic describes it as subtly biasing Claude's word choices in a way that doesn't change the meaning, quality, or readability of the response, but becomes statistically detectable over enough text. It travels with copied text and can survive light editing.
• Generated files get signed provenance metadata. SVG, PNG, and JPG files Claude produces carry metadata following the C2PA open standard, the same provenance system OpenAI and Google are also adopting for image outputs. This metadata signals Claude's involvement and can reveal tampering.

What this covers, and what it doesn't

Marking applies across Claude Platform (the API), Claude.ai, Claude Code, Claude Cowork, and Claude Tag, plus cloud partner surfaces like AWS, Google Cloud, and Microsoft Foundry where supported. Models launched August 2, 2026 or later support marking from day one. Older models are still being retrofitted, so coverage isn't instant or universal yet.

There's no setting to disable it, on any plan.

What to actually do about it

For most people using Claude to draft or polish marketing content, practically nothing changes day to day. The watermark doesn't alter what you get back, and there's no indication Google Ads or Google Search treats AI-assisted content differently just because it can technically be detected as AI-touched, quality and usefulness of the content is still what matters for ranking and ad approval.

Where it's worth a second thought: content you're publishing as explicitly human-authored, a founder's letter, a testimonial-adjacent piece, anything where "written by a person" is doing real work. If Claude did meaningful drafting on something like that, know that the underlying text carries a detectable signal now, even after you've edited it.

Where I'd push back a little

The stated goal, transparency about AI involvement, is reasonable on its face. The execution has drawn real pushback, and not without cause. There's no opt-out, which means someone who used Claude for a single early draft and then rewrote the piece entirely by hand still has a document that may register as AI-touched. Anthropic hasn't published much technical detail on how easy the text watermark is to detect or how reliable it is against false positives, which makes it hard for anyone to evaluate the actual risk of a piece of genuinely human work getting flagged. "Detectable over enough text" is also doing a lot of work in that sentence, and Anthropic hasn't said what "enough" means in practice.

None of that makes the feature illegitimate. It does mean treating "AI-detectable" as a precise, reliable signal right now would be overconfident, on Anthropic's part and on anyone else's.

Frequently asked questions

Can I turn off watermarking in Claude? No. It applies automatically at the model level across Claude.ai, the API, Claude Code, Claude Cowork, and Claude Tag, with no opt-out on any plan.

Does this affect content I've already published? Only if it was generated by a model that supports marking. Content from older, not-yet-updated models isn't retroactively marked, and Anthropic hasn't published a way to check existing content after the fact.

Will Google penalize ad copy or landing pages that Claude helped write? There's no indication of that. Google's public guidance has consistently focused on content quality and usefulness, not how it was produced. This watermark is about provenance disclosure, not a ranking or ads-policy signal.

Does editing AI-generated text remove the watermark? Not reliably. Anthropic says the watermark can persist through copying and light editing. Heavy rewriting may reduce or eliminate the detectable signal, but there's no published threshold for how much editing that takes.

Does this apply outside the EU? Yes. Even though the requirement originates from the EU AI Act, Anthropic applies the marking worldwide rather than geofencing it to European users.

Are other AI companies doing something similar? Yes. OpenAI joined the C2PA coalition and partnered with Google to embed Google's SynthID watermark into its image outputs. Adoption of provenance standards like C2PA is growing across the industry, though it's still inconsistent between providers.

Anthropic's own explanation is at How Claude marks AI-generated content.