Story
August 17, 2026
Claude’s Invisible Watermark Could Expose AI Cheating—and Ordinary Editing
Anthropic is preparing global watermarks for Claude text and files under EU transparency rules. Supporters see a needed check on hidden AI use; critics fear a blunt signal that can stigmatize human work merely touched by a chatbot.
Anthropic’s push to make Claude leave an invisible trace promises a cleaner line between human and machine-made work. But its watermark may create a more awkward distinction: between work written by AI and work merely processed by it.
The shift follows the EU AI Act’s new transparency obligations, which took effect on August 2. Anthropic says new supported Claude models will mark outputs from launch worldwide, while older models will be updated later. Text will receive an imperceptible watermark; supported files, including images, will carry C2PA provenance metadata.1
The company says the text system, based on Google DeepMind’s SynthID-Text approach, works by subtly steering low-stakes word choices into a pattern readable with the right key—not by inserting hidden characters or changing what readers see.2 Anthropic plans a detection API for users and third parties, answering concerns that it alone could become “judge, jury, and prosecutor.”3
Yet the rollout quickly exposed the policy’s central tension. A detected mark means Claude processed material, not necessarily that it authored it. Proofreading, translating, formatting or summarizing human work can leave a signal—potentially branding a lightly assisted press release, student paper or piece of code as AI-tainted.4 Critics argue that goes beyond the Act’s exemptions for routine editing and risks teaching teachers, publishers and employers to read “AI touched this” as “AI wrote this.”5
Anthropic also concedes the technology is no lie detector. Heavy rewriting, translation, short passages or mixed text can erase or weaken detection; absent a mark, meanwhile, does not prove a work is human-made.6 Supporters counter that imperfect provenance is still preferable to a world in which AI ghostwriting remains invisible, and could help curb AI-generated material feeding back into future models.7
For now, the watermark is best understood as a transparency signal, not a verdict. Its real test will come when institutions decide whether they can resist turning that signal into a scarlet letter.