Claude's text just got
an invisible watermark.
Anthropic says it will embed a statistical watermark into text generated by Claude. It's designed to survive copy-and-paste, and the direct trigger is compliance with the EU's AI transparency rules.
Not bolted on after —
woven into generation itself
Image watermarking has existed for a while. This pushes into the hardest domain to detect: plain text.
Anthropic announced that Claude models released after August 2, 2026 will automatically embed a statistical watermark into the text they generate. The mark survives copy-and-paste, letting anyone later verify whether a piece of writing was AI-produced. Per TechCrunch's reporting, this isn't a post-hoc pass applied after the text is written — it's a statistical bias woven directly into how the model selects its next token during generation.
The direct trigger is regulatory. Per Interesting Engineering's reporting, Anthropic signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, which requires generative AI providers to mark their output in a machine-readable way. Anthropic already used the C2PA industry standard for images and files, but its text-side coverage had lagged.
That gap wasn't unique to Anthropic. Across the industry, when OpenAI announced adopting C2PA and SynthID for images back in May 2026, text-output watermarking was left for later by every major vendor. This announcement is best read as Anthropic closing that gap. It's also broad in scope: the watermark applies across Anthropic's whole product line — the Claude app, the API, the developer tool Claude Code, the collaboration tool Claude Cowork, and Claude Tag for Slack — suggesting this is a company-wide policy shift, not a single feature add.
Coverage, at a glance
Who it affects, and how
Education & publishing
Opens the door to technically checking whether a submission or manuscript was AI-written, even after copy-paste. Whether a public detection tool actually exists is a separate, open question.
Corporate compliance
For companies operating in the EU, this gives one more data point for checking a vendor's regulatory posture, and can be checked against your own AI-use disclosure policy.
Everyday users
Nothing changes about how Claude looks or feels to use day-to-day. The watermark is a statistical property, not something you'd notice by eye.
Not a mark on the text —
a trace in how it was made.
Watermarks aren't foolproof
As Forbes reported, the announcement has drawn pushback online. Statistical watermarks can lose reliability under heavy rewriting, translation into another language, or short excerpted quotes. "Survives copy-paste" describes a degree of robustness — not detectability against every possible edit.
There's also a practical access gap. Embedding the watermark is only half the story; if the tool to actually read it isn't publicly available, schools and publishers have limited means to verify anything themselves. This announcement covers embedding — who gets to detect it, and how, remains an open operational question.
What to watch next is whether Anthropic makes a detection tool broadly available, and how interoperability shakes out with other vendors' image and text watermarking schemes. If standards stay fragmented, cross-checking "which AI made this" will remain weak in practice, whatever any single company does.