Anthropic has integrated a text watermarking feature into Claude's latest models to comply with the EU AI Act. The watermark is invisible to human readers but detectable by a proprietary scanner restricted to certain organizations.
TL;DR
- Anthropic's Claude now includes an EU-compliant AI watermarking feature in its latest models.
- The watermark is undetectable to human readers but can be identified by a proprietary scanner available to select organizations.
- This development has implications for AI transparency, hiring practices, and evidence verification.
What happened
Anthropic announced that it is watermarking all text produced by Claude to comply with the new EU AI Act. The watermark is integrated into Claude's newer models, with implementation for older models on the way.
The watermark aids in the detection of Claude-generated text but is indistinguishable to human readers, according to a company release. AI researcher Mustafa Ocal praises the approach, noting its complexity and human undetectability.
The watermark is only readable by a proprietary scanner provided by Anthropic. Currently in private preview, the scanner is available to a limited number of organizations, including law enforcement, regulators, media organizations, fact-checkers, and select researchers and businesses.
Why it matters
This development is significant for AI transparency, as it allows for the detection of AI-generated text without altering the user experience. It has practical implications for hiring practices, where companies may need to disclose AI involvement in decision-making processes.
The watermarking technique could also be valuable in rooting out forged documents and evidence. For example, the U.S. Department of Justice could use the scanner to verify the authenticity of evidence.
However, the limited access to the proprietary scanner means that the general public and many organizations will not be able to detect Claude-generated text. This raises questions about the broader impact and accessibility of the technology.
Key facts
- Anthropic's Claude now includes a text watermarking feature to comply with the EU AI Act.
- The watermark is integrated into Claude's newer models, with older models to follow.
- The watermark is undetectable to human readers but can be identified by a proprietary scanner.
- The scanner is currently in private preview and available to select organizations.
- The watermark is based on word choice and is practically invisible to human readers.
- The watermark is designed to be difficult to remove, requiring heavy editing or rewriting.
- Anthropic's approach is considered more comprehensive than Google's earlier implementation of text watermarking.
Context
Text watermarking is not a new concept, with Google pioneering the technique in 2024. However, Anthropic's approach is notable for its universal implementation across its latest models and the provision of a proprietary scanner for detection.
The EU AI Act requires AI systems to meet certain transparency and accountability standards. Anthropic's watermarking feature is a step towards complying with these regulations.
The development of AI detection methods is ongoing, with commercial AI detectors focusing on differences in AI and human writing styles. Anthropic's watermarking technique, however, is based on word choice and is designed to be indistinguishable to human readers.
