Anthropic Will Now Watermark Every Word Claude Writes

Every piece of text Claude generates from now on carries something you can’t see: an invisible marker proving where it came from. Anthropic confirmed this week that it’s watermarking AI-generated text and files across its Claude products, a move driven less by choice than by a regulatory deadline the company had no way around.

The confirmation came through an updated Anthropic support page, first reported by TechCrunch. It ties directly to the European Union’s AI Act Transparency Code, which took effect on August 2 and requires AI companies to mark generated or edited content in ways other systems can detect.

What Anthropic is actually doing

Anthropic’s approach splits into two tracks, depending on what Claude produces. Generated text gets an embedded watermark woven directly into the words themselves, invisible to anyone reading it, but detectable by systems built to check for it. Generated files, meaning images and documents in formats Claude supports, get signed provenance metadata following C2PA, the open standard already used across the industry for tracking content origin.

The text watermark is the more unusual piece of engineering here. According to Anthropic, it survives copy-paste. Someone can lift Claude’s output out of the chat window and drop it into an email or a CMS, and the mark travels with it. The company says the watermark also “may persist through some editing”, though it hasn’t specified how much editing breaks it.

Coverage extends across the board: Claude Platform (the API), the Claude consumer app, Claude Code, Claude Cowork, and Claude Tag all carry the same watermarking, since Anthropic applies it at the model level rather than per product. That means the mark doesn’t depend on which interface someone used to reach Claude. It’s part of the model’s output itself.

The rule applies fully to any Claude model launched on or after August 2, 2026. Older models aren’t covered yet. Anthropic says it’s working to extend support backward but hasn’t given a timeline for when every previous model will carry the same watermarking.

Why now

The EU’s Transparency Code isn’t optional guidance. It’s a compliance mechanism tied to the AI Act, and Anthropic is one of several major AI companies that have already committed to following it, alongside Black Forest Labs, Google, Meta, Microsoft, OpenAI, and Synthesia. Missing the August 2 deadline would have put Anthropic on the wrong side of enforceable EU rules in a market it can’t afford to exit.

But the regulatory push is landing at the same moment platforms are moving on their own, for reasons that have nothing to do with Brussels. Suno said last week it will start marking AI-generated songs, a move that came after a string of legal disputes over how the platform’s music gets used and attributed. Substack partnered with detection tool Pangram last month to flag AI-written newsletters, after its CEO Chris Best publicly criticized what he called “Claudefishing”, writers passing off AI-generated newsletters as their own work without disclosure.

Read together, these moves suggest watermarking is becoming table stakes across the AI industry, not a differentiator. Regulatory pressure from the EU is forcing the issue for some companies. Reputational pressure from users tired of undisclosed AI content is forcing it for others.

What this means for users

For most people using Claude to draft emails, summarize documents, or write code, nothing changes about the experience. The watermark is invisible, doesn’t affect output quality, and doesn’t require any action to enable. Anthropic has been explicit that it doesn’t alter meaning or readability.

Where it matters is downstream. Anyone publishing Claude-generated text, whether that’s a journalist, a student, a marketer, or a content creator, should now assume that text carries a detectable signature, even after it’s been copied elsewhere. Anthropic says it’s building tools to let users and third parties check for that signature, though the technical details of how detection will work haven’t been published yet.

There’s a real gap in what Anthropic has confirmed versus what it hasn’t. The company won’t say how much editing removes the watermark. That’s not a small omission. For platforms like Substack that are actively trying to catch undisclosed AI writing, the answer to “how much rewriting defeats detection” determines whether this watermark is a meaningful transparency tool or a marker that vanishes the moment someone runs a paraphrase pass.

What this means for the industry

For AI companies broadly, this locks in a new baseline. Once Google, Meta, Microsoft, OpenAI, and Anthropic have all committed to the same EU code, watermarking stops being a feature any one company can use as a selling point and becomes closer to a compliance cost every major model provider has to absorb. The competitive question shifts from “Does this company watermark its output?” to “How good is the detection, and how hard is it to strip?”

That second question is where the real industry pressure will show up next. Detection tools like Pangram are already being deployed by publishers specifically to catch AI writing that platforms themselves aren’t flagging. If Anthropic’s watermark turns out to survive only light editing, third-party detectors and Anthropic’s own official detection method could start disagreeing with each other, which would undercut the credibility of watermarking as a transparency mechanism just as it’s becoming standard.