Anthropic's Claude watermark goes further than EU law requires

Anthropic's new watermark for AI-generated text has drawn sharp reactions from Claude users. A Business Insider review of the relevant EU law, backed by interviews with policy experts, suggests the company's approach reaches beyond what the regulation demands.

Article 50(2) of the European Union's AI Act does not require AI companies to mark outputs when the technology is used for standard editing or does not substantially change a user's original words or meaning. Anthropic's watermark is designed to survive copy-and-paste. It can appear in translations, summaries, and heavily edited text.

"Article 50 called for a scalpel to catch deepfakes and mass synthetic generation," Caroline De Cock, senior fellow at the Lisbon Council think tank and head of research at the Information Labs, told Business Insider. "Anthropic is using a sledgehammer that hits simple proofreading and translation."

Anthropic has said it is introducing the watermark to comply with the EU's AI Act. The company is rolling it out globally because it cannot apply it on a per-region basis.

The announcement triggered a wave of backlash. Some Claude users said they canceled their subscriptions. Developers have created tools to remove the watermark.

A recurring concern is that the watermark could be interpreted as indicating that Claude wrote the text, even if it was used only for editing or proofreading. The technical approach behind the watermark guides Claude's word choices to create a detectable pattern. That pattern signals the chatbot was likely involved in writing the text.

Lighter editing with Claude is less likely to trigger the watermark, an Anthropic spokesperson told Business Insider, because "there's very little (if anything) for the watermark to attach to."

Milos Rusic, the cofounder and CEO of enterprise AI company Deepset, told Business Insider that Anthropic "is taking a somewhat broader interpretation" of the EU law. A "provenance signal" can attach to Claude's word choices even when the model "did not originate the underlying idea or document," he said.

The approach fits Anthropic's "brand identity as a safety-first AI developer," De Cock said. It could also reduce the company's regulatory risks by "shifting the burden of filtering and interpreting those labels entirely onto downstream users and external platforms."

An Anthropic spokesperson told Business Insider the company would "continue to evaluate different approaches."

The EU's watermarking laws apply to new models launched in Europe on or after August 2. Older models have a grace period until December. Other tech companies operating on the continent will have to roll out their own machine-readable ways to detect AI-generated content in the coming months.

Google said it uses its SynthID technology to invisibly watermark AI-generated text, images, audio, and video from its models worldwide. OpenAI referred Business Insider to an FAQ saying it plans to expand its labels to text without specifying a date. Meta directed Business Insider to a July blog post about its work on practical, interoperable tools to identify AI-generated content. Meta has focused on images so far and has not provided details on a text-watermarking plan.

xAI did not respond to a request for comment. The company tags some work as AI-generated, including White House press secretary Karoline Leavitt's departure message on X. The tag later disappeared.

The coming wave of watermarks arrives as many companies encourage employees to use AI. Some people now worry about being outed and having their authorship questioned.

"These systems can be useful evidence," Rusic said of watermarks. "I would be very cautious about turning them into an automated judgment about authorship or misconduct."