Claude Now Watermarks Its Text — and Detection Is Messier Than It Sounds

Somewhere between "AI detection is solved" and "watermarks are pointless" sits the actual news: Anthropic has started embedding invisible watermarks in Claude's text output, everywhere in the world, and both the hopes and the fears around it are overdone.

What happened

For models launched in the EU after August 2, Anthropic is marking AI-generated content in two ways, according to Axios: text watermarks — patterns embedded in the generated text that Anthropic describes as imperceptible — and file metadata, digital signatures attached to media files Claude processes. The stated driver is Article 50 of the EU AI Act, which mandates disclosure of AI-generated or manipulated content. Anthropic is applying the marks wherever Claude is offered, worldwide, not just in Europe. Providers have until December 2 to bring legacy models into compliance, so expect other labs to follow: OpenAI has already outlined its own compliance approach, though its watermarking currently focuses on images and audio rather than text.

Anthropic itself flags two significant limitations. First, AI-assisted can look like AI-generated: a press release written by a human but proofread, formatted, or translated by Claude can carry the mark. Second, detection drops off when text is heavily rewritten, mixed with other copy, or too short.

What it means for how you work

If your team uses Claude to polish human-written content — and most comms teams I know do — the risk is real but specific: documents that were never "AI-written" could start registering as AI-flagged, with everything that implies for client deliverables, journalism workflows, and internal trust. If you're a student or a professional whose workplace is about to buy AI-detection tooling, know that the detector's word is not final; both false positives and deliberate evasion are built into the system's design constraints.

The practical move is policy, not panic: be clear internally about where AI assists versus authors, because the watermark can't tell the difference and your processes need to.

My take

My honest POV is a practicality question: how enforceable is this really? Anyone with the knowledge and determination can get rid of most of these watermarks. This isn't a crackpot take — Theo (t3.gg) has an extended hands-on video walking through exactly how removable these schemes are in practice, and it's worth watching before you treat detection as ground truth: https://youtu.be/Be-NqsW-wuk.

So the watermark will catch the careless, not the committed. That's still not nothing — casual misuse, the student who pastes an essay wholesale, the vendor quietly shipping raw model output — is exactly the tier that gets caught. But as compliance infrastructure, it's soft: a speed bump for the lazy, invisible to anyone motivated. The EU-mandated disclosure regime is coming regardless, so the smart play is to build honest AI-usage habits now, before a December deadline and a wave of detector vendors make the question uncomfortable.