Anthropic Embeds Watermarks in Claude to Track AI-Generated Text

Anthropic introduces text watermarking across Claude models

Anthropic announced Monday that its Claude AI models will now embed invisible watermarks directly into generated text. The watermarks remain intact when content is copied and pasted, and may survive some editing, according to the company.

The feature launches as part of Anthropic's compliance with the European Union AI Act's transparency requirements. All Claude models released on or after August 2 include watermarking from launch, with the company working to add the capability to earlier versions. The watermarks apply globally to all Claude-generated content, including output from the AI when accessed through cloud service providers.

According to Business Insider, which first reported the details, Anthropic plans to provide third-party tools that can detect these watermarks, potentially giving publishers, educators, and other organizations a new method to identify AI-generated writing.

Why it matters

The publishing industry has struggled with AI-generated content scandals that damage author reputations and publisher credibility. Watermarking technology offers a technical solution to a problem that has relied on subjective judgment and circumstantial evidence. For educational institutions facing widespread AI use in student work, detection tools could help enforce academic integrity policies. However, the technology's limitations mean it functions as one verification method among many, not a definitive solution.

Recent controversies highlight the need

The publishing world has seen several high-profile cases involving suspected AI writing. Last month, a literary agent withdrew support for the crime novel "Call Me, I'll Hide the Body" after AI use concerns emerged, despite fourteen publishers bidding for rights and a deal already closing. Author Jerry Falade denied using AI.

Earlier this year, Hachette pulled Mia Ballard's horror novel "Shy Girl" following AI-generated writing allegations. Ballard told The New York Times she hadn't used AI to write the book, but that a freelance editor had introduced AI-generated material without her knowledge.

Significant limitations remain

Anthropic acknowledged that its watermarking system has weaknesses. Heavy editing, paraphrasing, translation, or combining Claude's output with other writing can render the watermark undetectable. Additionally, the presence of a watermark doesn't definitively prove Claude authored the original content—even using the AI for proofreading or translation can leave a detectable mark.

These limitations mean watermarks serve as indicators rather than proof, requiring human judgment to interpret results in context.

Second major lab to deploy text watermarking

Anthropic follows Google DeepMind in implementing text watermarking. In 2024, Google announced it was watermarking text and video content generated through the Gemini app and web interface using its SynthID technology, building on an image watermarking feature released in 2023.

The details were first reported by Business Insider.