Anthropic Unveils Invisible Watermark for AI Texts

⚡

Key Takeaways

1Anthropic launches an invisible watermark to identify texts generated by Claude AI, complying with the EU AI Act.

2This watermark, embedded in the text, does not alter its meaning and persists despite certain edits, but can be removed through heavy modifications.

3The initiative aims to assist publishers and institutions in detecting AI-generated content, following controversies in the literary world.

💡Why it matters — This technology could transform how AI-generated content is verified, impacting the publishing industry and education.

Anthropic Introduces a Watermark for AI Texts

Anthropic, a company specializing in artificial intelligence, has recently unveiled a new feature aimed at enhancing the transparency of content generated by its AI models. This initiative aligns with a commitment to comply with European legislation on artificial intelligence.

An Invisible Watermark for Claude AI

The company announced the integration of invisible watermarks in the texts produced by its Claude AI model. This mechanism, although imperceptible to the naked eye, is designed to persist even when the text is copied or pasted. However, it is important to note that heavy editing can erase this watermark, and its presence does not necessarily guarantee that the text was generated by an AI.

This innovation will make it more difficult to use Claude AI as a ghostwriter for works such as novels or school assignments. The watermark, while invisible, accompanies the text through its various transformations without altering its readability or meaning.

Compliance with European Legislation

This measure is part of Anthropic's commitments to the EU AI Act. The Claude models, launched starting August 2, incorporate this feature from their inception. Anthropic is also working on extending this technology to earlier versions of its models. The marking applies universally, even when Claude is accessed via cloud services. Furthermore, Anthropic plans to provide tools to third parties to detect these watermarks.

Impact on the Publishing Industry

The watermarking could have significant repercussions for the publishing industry, which is currently grappling with challenges posed by AI-generated content. Recently, a literary agent withdrew support for a bestselling novel, "Call Me, I'll Hide the Body," after suspicions arose regarding the author's use of AI. Jerry Falade, the author, denied these allegations.

Similarly, Hachette pulled a horror novel by Mia Ballard, titled "Shy Girl," following similar accusations. Ballard claimed that a freelance editor had introduced AI-generated content without her consent.

Limitations and Challenges of the Technology

Despite its advantages, Anthropic's watermark has its limitations. Heavy editing, paraphrasing, translation, or mixing with other texts can render the watermark undetectable. Moreover, the presence of a watermark does not prove that Claude originally wrote the text, as even proofreading or translation by AI can leave a trace.

Anthropic is not the only company exploring this avenue. In 2024, Google DeepMind also announced the use of its SynthID technology to mark texts and videos generated by its Gemini application, furthering its efforts to enhance the traceability of AI content.