It’s going to get a lot harder to let Claude be a ghostwriter for everything from novels to homework.
On Monday, Anthropic said new Claude models will embed an “imperceptible watermark” directly into AI-generated text. The watermark doesn’t change the text’s meaning or readability, but travels with it when copied and pasted and “may persist through some editing,” the AI lab said.
A watermark is a faint pattern or text embedded into photos or digital documents to help prove ownership and stop people from copying or stealing a piece of work.
The move is part of Anthropic’s commitments toward more transparency under the European Union AI Act. Claude models launched on or after August 2 will support marking from launch, and Anthropic says it’s working on adding the function to older models. The markings apply to content produced by Claude worldwide, including when Claude is accessed through cloud providers. Anthropic said it also plans to provide third parties with tools to detect them.
AI watermarking could be a turning point for the publishing industry, which has been grappling with AI-generated content.
Last month, a book agent pulled support for the hot crime novel “Call Me, I’ll Hide the Body” after concerns were raised about the author possibly using AI. Fourteen publishers bid for the book’s publishing rights, and a deal was closed before the agent apologized. The writer, Jerry Falade, denied using AI to write the book.
Earlier this year, Hachette pulled Mia Ballard’s horror novel “Shy Girl” after allegations of AI-generated writing. Ballard told The New York Times she hadn’t used AI to write it, adding that a freelance editor had introduced AI-generated material without her direct knowledge.
Claude’s new watermark could give publishers, schools, and universities another way to investigate such cases and make it harder for people to pass off AI-generated writing as their own.
Still, there are loopholes. Anthropic said that heavy editing, paraphrasing, translating, or mixing Claude’s output with other writing can make its watermark undetectable. Plus, finding a watermark doesn’t prove Claude originally authored something, since even proofreading or translating with its AI can leave a mark.
Anthropic is the second major lab to introduce text watermarking.
In 2024, Google DeepMind announced it was watermarking text and videos generated in the Gemini app and web experience using its SynthID technology. It had released an image version of the feature in 2023.
Read the full article here
