Anthropic Claims to Mark AI Text but Won't Say How

Anthropic published a support page on how Claude marks AI-generated content, announcing that it has signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content as both a generative AI model provider and generative AI system provider. The company states it is 'planning to put those commitments into practice' and will 'update this article and publish more detailed technical guidance as it becomes available.'

For plain text output, Anthropic says that when a supported Claude model generates text, it 'weaves an imperceptible watermark directly into the text itself.' The company claims users 'won't see it' and that it 'doesn't change the meaning, quality, or readability of Claude's response.' Anthropic adds that because the watermark is part of the text, it 'will travel with the text when it's copied and pasted elsewhere, and may persist through some editing.' The watermarking will be applied at the model level, meaning it will be present 'no matter which Claude product or surface the text comes from.'

Anthropric is also 'working to enable users and other third parties to detect Claude's embedded watermarks and provenance metadata.' According to the company, detection 'checks whether a piece of text or a file carries a supported Claude mark,' and 'if a supported mark is found, it indicates that the content may have been processed by Claude.' The company states it will 'share details on detection mechanisms in forthcoming technical documentation.'

Anthropric notes that marking will apply 'to output from supported models wherever Claude is offered, worldwide,' despite the commitment being motivated by EU regulation. However, Anthropic has not disclosed the specific technical method used to create the watermark—whether it relies on invisible characters embedded between visible text, statistical word-choice patterns (steganography), or another approach. This lack of explanation has prompted scrutiny over practical concerns, including whether invisible character sequences could inflate character counts on social media platforms with length limits, whether statistical watermarking could force suboptimal word selection and degrade output quality, and whether quoted or copied text could inadvertently carry the watermark into users' own writing and falsely mark it as AI-generated. Anthropic has not addressed these questions in its initial announcement.