Anthropic is adding invisible watermarks to text generated by supported Claude AI models to help comply with European Union artificial intelligence regulations. The company said similar measures are being adopted by other AI laboratories as regulators push for clearer identification of AI-generated content.
“We are adding markings to Claude’s output to comply with EU AI laws, and other laboratories are taking similar steps,” Anthropic said. “AI-generated text is difficult to identify, but this provides better tools for identification. Text from supported Claude models, including output generated with Claude Code, will contain an invisible watermark. The meaning, quality, and readability of Claude’s responses will remain unchanged. We also plan to launch a text-detection API so users can conduct additional checks themselves.”
EU AI transparency rules increase pressure on AI companies
The European Union’s AI transparency requirements are designed to ensure that systems such as Claude do not “undermine the integrity and trust of the information ecosystem” or create new risks, including misinformation, large-scale manipulation, fraud, identity theft, and consumer deception. Article 50 of the EU AI Act is expected to become a major compliance challenge for generative AI companies.
“People should be aware when they are interacting with AI or being exposed to content generated by AI,” the European Commission said in its AI transparency guidelines. “This can help you make informed decisions, calibrate your trust and reliance on AI, and avoid misinformation and deception.”
However, the rules raise questions about how AI-generated content should be disclosed to the public. For example, an article initially written with Claude but substantially reviewed and edited by a human could still contain an invisible watermark without being visibly labeled for readers.
The EU expects AI developers such as Anthropic to play a central role in creating reliable watermarking systems. As generative AI continues to advance, regulators say there will be a growing need for technologies that can trace the origin of digital information.
At the same time, the practical value of AI watermarks will depend largely on how technology companies implement them. EU rules require watermarking and other identification methods to be technically feasible, sufficiently reliable, interoperable, effective, and resistant to manipulation.
Anthropic said it will continue developing AI watermarking and text-detection tools to meet EU requirements. Failure to comply with the EU AI Act’s penalties could result in fines of up to €15 million or 3 percent of a company’s global annual revenue.
Ars Editor-in-Chief Ken Fisher contributed to this report. This story was updated on Aug. 13 to include a statement from Anthropic.
Source: arstechnica.com