Anthropic starts watermarking all new Claude outputs to meet AI transparency rules

Anthropic is rolling out a major change to how its Claude models leave a trace on the internet. Starting with models launched on or after August 2, all newly generated Claude text carries an invisible, statistically detectable watermark. The scheme doesn’t rely on hidden characters or formatting; instead, it slightly biases token choices in ways that can later be verified using a public key, allowing a checker to say how likely it is that a given passage was produced by Claude.

The move is widely seen as an attempt to get ahead of the EU AI Act’s transparency provisions, which require clearer labeling of AI‑generated and manipulated content. Importantly, this is not a generic “AI detector”: it only answers “did Claude write this?”, and even then the signal breaks down if another model paraphrases or heavily edits the text. That nuance has already sparked debate. Educators and compliance teams may be tempted to treat watermark hits as proof of cheating, while skeptics warn of false accusations and over‑reach. At the same time, publishers and platforms drowning in low‑quality AI spam hope vendor‑specific watermarks will give them a practical tool to filter out bulk‑generated content. How regulators, institutions, and platforms actually use these signals will shape the real impact of Anthropic’s decision.