Anthropic launched an AI text watermark that remains detectable after copying and pasting, and the move has triggered subscriber cancellations along with new tools designed to strip the mark away. Yet legal experts say the European Union never demanded a system this aggressive.
The backlash matters because companies across Europe now face the same regulatory deadline. Under Article 50(2) of the EU AI Act, providers must label machine-generated output, but the requirement does not apply when AI assists with standard editing or when it leaves a user’s original wording and meaning largely intact.
Anthropic’s watermark, however, works differently. It steers Claude’s vocabulary choices to create a detectable pattern, which means the mark can appear even in translations, summaries, and heavily revised drafts. The company said it rolled the feature out worldwide because it lacks the ability to restrict it by region.
“Article 50 called for a scalpel to catch deepfakes and mass synthetic generation, but Anthropic is applying a sledgehammer that hits simple proofreading and translation,” said Caroline De Cock, senior fellow at the Lisbon Council think tank and head of research at the Information Labs.
As a result, a persistent worry has emerged. The watermark could imply Claude authored a document even when the tool only cleaned up grammar or tightened phrasing. Milos Rusic, cofounder and CEO of enterprise AI firm Deepset, described Anthropic’s stance as “a somewhat broader interpretation” of the law. A provenance signal can attach to word choices, he noted, even when the model “did not originate the underlying idea or document.”
De Cock suggested the approach aligns with Anthropic’s safety-first brand identity. It may also lower regulatory exposure by pushing the burden of label interpretation onto users and platforms. An Anthropic spokesperson told Business Insider the company would “continue to evaluate different approaches,” adding that lighter edits rarely trigger the mark because there is little for the watermark “to attach to.”
Other tech giants are now scrambling to comply. New models launched in Europe after August 2 require machine-readable detection, while older models have until December. Google already uses SynthID across text, images, audio, and video. OpenAI said it plans to extend its labels to text without specifying a date. Meta pointed to a July blog post about interoperability but offered no text-specific timeline. xAI did not respond.
Meanwhile, employees who embraced AI at work now fear their authorship will face scrutiny. Rusic cautioned against treating watermarks as proof: “I would be very cautious about turning them into an automated judgment about authorship or misconduct.”
