Anthropic has added an invisible watermark to text generated by its Claude AI models to allow computer systems to detect AI-generated content [1, 2].
The move creates a new technical barrier for users attempting to pass off AI-generated writing as human-authored work in professional or academic settings. While the watermark remains imperceptible to human readers, it provides a digital trail that can be verified by software [1].
Anthropic said new Claude models will embed an "imperceptible watermark" directly into AI‑generated text [1]. This technical implementation ensures that the content is identifiable to computer systems without altering the visual appearance of the text for the end user [1, 2].
The company rolled out this policy to satisfy the EU AI Act's Transparency Code [2]. This regulation requires tech companies to label content that has been AI‑generated or edited in a manner identifiable to computer systems [2].
Reports indicate the rollout occurred earlier this month [1, 2]. The feature applies globally to the online Claude AI platform [2].
The introduction of the watermark has drawn attention from users who previously relied on the chatbot to complete assignments or workplace tasks undetected [2]. Because the watermark is embedded in the structure of the output, it differs from visible labels or disclosures that a user could simply delete before submission [1].
Anthropic said the policy is a necessary step to meet the legal requirements of the EU's transparency framework [2].
“Anthropic said new Claude models will embed an "imperceptible watermark" directly into AI‑generated text.”
This development signals a shift from voluntary AI disclosure to technical enforcement. By embedding detection capabilities directly into the model's output, Anthropic is moving toward a system where AI provenance is baked into the data. This reduces the effectiveness of 'stealth' AI use and aligns the company with the EU's stringent regulatory approach to AI transparency, potentially setting a precedent for other major LLM providers.
