Anthropic says its Claude invisible watermark plan will place machine-readable marks in text generated by future models and attach digitally signed provenance metadata to supported generated files. The policy is set to apply globally, including Claude Code and Claude Cowork, rather than being restricted to Europe, according to Forbes.
The immediate practical consequence is limited: users will not be able to inspect or independently test the marks yet. Anthropic has said it plans to publish information on recognizing or detecting them later, Ars Technica reported. Until then, the company has not publicly established the implementation, detection accuracy or how reliably the marks persist.
How will Claude’s invisible watermark work?
For generated writing, the proposed watermark is embedded in the text and invisible to the user. Forbes reported that Anthropic says it should remain present when text is copied and pasted. Generated files, including images where supported, would instead carry signed provenance metadata, a record intended to indicate origin.
Neither method is durable in every circumstance. Anthropic has warned that substantial editing may erase a text watermark. Metadata can disappear when a file changes format, when someone takes a screenshot, or through other transformations, Forbes reported. A watermark can therefore provide a signal about a file or passage, rather than a permanent and conclusive history of every revision it has passed through.
Users cannot opt out under the announced policy, according to Forbes. That gives the decision relevance beyond people who use Claude to draft from a blank page: the policy covers all Claude models and products worldwide. But the exact boundary of what will be marked needs more work before anyone claims certainty.
Will Claude watermark proofreading and edited user text?
The reporting points in two directions. Ars Technica said Anthropic described a broader approach that would watermark content processed by its models where supported, not solely material the model created. Forbes and CNN characterized the program as marking Claude-generated content, with CNN saying future Claude models will watermark generated text.
That difference matters for people using Claude as an editor, proofreader or coding assistant. European rules described by Ars Technica exclude some standard editing assistance and changes that do not substantially alter a user’s text or its meaning. The available reporting does not settle whether Anthropic will mark lightly edited user writing, so users should not treat that as confirmed either way.
Why is Anthropic adding the marks?
Anthropic has tied the change to European Union AI Act transparency requirements, according to Ars Technica and CNN. Ars Technica reported that the rules require providers to mark AI-generated or manipulated audio, images, text and video, with a transition period through December 2026 for models released before the relevant deadline.
The announcement has already drawn objections online from people who say they use Claude to proofread their own work and developers worried about possible effects on code, Forbes reported. Those are user concerns, not demonstrated technical outcomes. The available evidence does not show whether the system changes text quality, readability, meaning or code output.
This story draws on original reporting from Ars Technica.
