Claude Could Put an Invisible AI Watermark on Writing You Wrote Yourself

Anthropic announced on August 10 that it will add invisible, machine-readable watermarks to text generated by supported Claude models, rolling out worldwide to comply with the European Union's voluntary transparency code under the AI Act. The watermark travels with copied text and may survive editing, but Anthropic cautions that its presence does not prove AI authorship, since human-written text processed by Claude (e.g., proofreading) can carry the mark, and heavy editing can remove it. Critics including attorney Peter Harrell and political scientist Maya Sen argue the system mislabels human writing that uses AI for cleanup, while Y Combinator co-founder Paul Graham floated a startup idea to rephrase AI output and 'obliterate the watermarks.'

You could write every word yourself, ask Claude to fix a few typos and still wind up with writing carrying an invisible AI watermark. That's one of the stranger catches in Anthropic's new plan to make AI-generated content easier to identify. In a Help Center article https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content published Monday, August 10, Anthropic detailed how it plans to add machine-readable markings to content generated by supported Claude models. The move comes after the company signed the European Union's voluntary transparency code https://digital-strategy.ec.europa.eu/en/policies/code-practice-ai-generated-content , which outlines ways providers can comply with the AI Act's rules on AI-generated content. The system won't be limited to Europe, though. Anthropic says the markings will roll out worldwide wherever Claude is offered. For text, Claude will weave what the company calls an "imperceptible watermark" directly into its responses. Users won't be able to see it, and Anthropic says it won't affect the meaning or readability of what Claude produces. Simply moving the text somewhere else won't necessarily get rid of it. Because the watermark is part of the text itself, Anthropic says "it will travel with the text when it's copied and pasted elsewhere" and may survive some editing. That sounds relatively straightforward until you get to one of Anthropic's own caveats: finding a watermark does not prove Claude actually wrote the material. Someone could hand Claude something they wrote themselves and ask it to proofread, translate, summarize or otherwise process the text, and the resulting version may still carry the mark. At the same time, Claude-generated writing can lose its detectable watermark after enough editing, paraphrasing, translation or mixing with other text. Very short passages may not contain enough signal for reliable detection at all. In other words, by Anthropic's own description, a watermark isn't a foolproof test of AI authorship — and the absence of one doesn't prove something was written entirely by a human. That's the distinction now drawing pushback. Attorney Peter Harrell, a visiting scholar at Georgetown Law's Institute of International Economic Law, wrote https://x.com/petereharrell/status/2087020627509100824 that if he understands the policy correctly, uploading his own writing for a copy edit means Claude will "watermark my human written, AI-copy edited text as AI, which seems ridiculous." Political scientist Maya Sen raised a similar concern https://x.com/maya sen/status/2086966088324047258 , saying she dictates a lot and uses AI to clean up transcribed text. The change, she said, would "make Claude a lot less useful for users like me." On Hacker News https://news.ycombinator.com/item?id=49250109 , commenters also pointed out that writers with dyslexia can rely on essentially the same workflow. Cryptographer Ian Miers argued https://x.com/secparam/status/2086988857057169493 the premise was already outdated. “In 2026, even the idea of ‘pure human text’ feels like a relic,” he wrote, adding that a watermark can't distinguish between a human using AI assistance and AI working on its own. Y Combinator co-founder Paul Graham, meanwhile, was already thinking about how to beat the system. Within hours, he floated a startup idea https://x.com/paulg/status/2086924675163345285 that would rephrase AI output while preserving its meaning and "obliterating the watermarks." Exactly how Anthropic's invisible text watermark works remains unclear. The company hasn't publicly detailed the underlying technique and says it is still developing ways for users and third parties to detect the markings, with more technical documentation to come. There's also an important timing wrinkle. Anthropic says models launched on or after August 2 will support marking from launch, while support for models released before that date remains in progress. Anthropic's currently listed Claude models all predate that cutoff, and the company hasn't said when the retrofit will be complete. Anthropic isn't alone in signing the voluntary code. OpenAI, Google, Meta, Microsoft and Mistral are also on the EU's current list of signatories https://digital-strategy.ec.europa.eu/en/news/strong-backing-code-practice-transparency-ai-generated-content , while xAI is not. That doesn't exempt non-signatories from the AI Act's underlying transparency rules, however. Companies covered by Article 50 https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-50 still have to comply and show that their own measures are adequate.