Anthropic’s Claude Gets Invisible Text Watermarks That Proofreading Alone Could Trigger

Anthropic is introducing machine-readable watermarks into text generated by its Claude artificial intelligence models, a move driven by new European Union rules that could reshape how AI-assisted content is created.

The AI company said models that have been launched in the European Union after Aug. 2 will mark Claude-generated content in two ways. The system will apply wherever those models are offered worldwide, meaning the changes will not be limited to European users.

For text, Claude will embed what Anthropic calls "imperceptible" patterns to indicate that the content was generated or processed by the model. Media files created or modified with Claude will also contain metadata with digital signatures that can verify the asset passed through Anthropic's system.

The changes are intended to help the company comply with Article 50 of the European Union's AI Act, which establishes transparency requirements for providers of generative AI systems and requires certain AI-generated or manipulated content to be identifiable.

However, a watermark does not necessarily mean AI wrote the underlying content. Anthropic acknowledged that text could carry a detectable watermark even when Claude's role was limited to proofreading, formatting or translating material originally written by a person.

That distinction could be particularly important for news organizations, public relations firms, corporate communications departments and other businesses that increasingly use generative AI as an editing tool rather than as an author.

The watermarking technology also has limits in the opposite direction. Anthropic says detection can become less reliable when Claude-generated text is substantially rewritten, combined with other material, or is too short.

Earlier attempts often relied on outside tools that analyzed characteristics such as word choice, sentence structure and predictability to estimate whether a person or a machine wrote something. Watermarking instead puts part of the responsibility on AI developers to build signals directly into their models' outputs.

Anthropic is not alone in preparing for the EU's transparency requirements. OpenAI has also outlined its approach to complying with the AI Act, although its current watermarking and provenance efforts have focused primarily on images and audio rather than text.

Online platforms are simultaneously stepping up efforts to distinguish authentic material from mass-produced AI content. LinkedIn has been testing a "seems like AI slop" reporting option, while Substack has integrated technology from AI detection company Pangram. Snap has also stopped promoting AI-generated videos in its main Spotlight feed. AI companies have until Dec. 2 to bring older models into compliance with the EU requirements.