Anthropic has provided further insight into how it will implement watermarking in text generated by its AI chatbot, Claude. This move follows the company’s announcement earlier this month that watermarking would be added to comply with the EU AI Act’s Transparency Code, which mandates AI-generated content be identifiable. The watermarking will create subtle, undetectable patterns in Claude’s output by making “low-stakes choices” in word selection, such as choosing between synonyms. The company emphasized that the watermark will not affect the quality or readability of the text, making watermarked and unwatermarked responses indistinguishable to users.

Anthropic plans to utilize the SynthID Text watermarking method, originally developed by Google DeepMind in 2024, and will offer a detection API to verify if content is watermarked. This approach differs from typical AI content detectors that look for stylistic clues or common phrases indicative of AI, instead embedding a coded signal in the text itself. The company acknowledged that light editing might reduce the watermark’s visibility but won’t fully remove it, while more extensive rewrites where most words are changed can effectively erase the watermark, raising questions about whether the altered text can still be considered AI-generated.

Regarding AI-assisted editing and proofreading, Anthropic noted that if Claude only partially edits a human-written text, the watermark would be minimal or absent because few words originate from Claude’s generation. When it comes to code, the watermark’s presence will be limited as the AI must generate precise syntax without flexibility. However, in areas where wording choices are optional, such as in comments within the code, the watermark may still appear. Anthropic stresses that watermarking will have an insignificant effect on the functional aspects of code.

Anthropic also clarified that Claude will not be the only model embedding watermarks, as other major AI companies bound by the EU’s Code of Practice are expected to implement similar measures. While the move aims to enhance transparency and compliance with regulations, it has sparked some backlash among users, with reports of subscription cancellations linked to concerns over watermarking. Despite mixed reactions, this development marks a key step toward standardized AI content identification in the industry.