Anthropic's invisible Claude watermarks are live, and some users are not happy
Every Claude model released since August 2 now embeds an invisible, machine-readable watermark in text and C2PA metadata in images. The rollout reaches the API, Claude Code, and major cloud hosts, and some paying users say it will expose them.
Anthropic's invisible watermarks for Claude are no longer a policy commitment. They are shipping. Every Claude model released on or after August 2 now weaves a machine-readable signal into the text it generates, Fortune reported Monday, and images carry digitally signed provenance metadata under the open C2PA standard. Because the watermark operates at the model level, it follows Claude everywhere the model runs: the consumer chatbot, the API, Claude Code, and third-party hosts.
GaaS News covered the policy mechanics last week, when the move surfaced as part of Anthropic's commitments under the EU AI Act's transparency rules, which took effect August 2 and require providers of generative models to make synthetic output machine-readable and detectable. The follow-through is the news. Plenty of labs have promised provenance tooling. Anthropic has now turned it on by default, for everything, including sessions where Claude merely edits or translates text a human wrote.
The technical fine print matters for anyone reselling Claude output as a service. According to Fortune's account, the signal is designed to survive copy and paste and some light editing, while a heavy rewrite or a translation may remove it. Anthropic also cautions that the mark "only shows Claude had a hand in something, not that a Claude model generated the entire thing." That distinction will do real work in disputes over agency deliverables, ghostwritten reports, and agent-generated code comments.
Not everyone is celebrating. TechCrunch reported Wednesday that some users are openly angry, arguing the marks will flag Claude use to employers and professors. One critic on Reddit described the result as a "digital tattoo" on anyone who let Claude reorganize a paragraph. Most commenters in the threads pushed back, siding with detectability over deniability, but the complaint identifies the constituency with the most to lose: individual subscribers who built quiet Claude workflows into jobs and coursework that forbid them.
For agentic AI vendors the open questions sit on the detection side. Anthropic has not published details on who gets access to the detector, what its error rates look like, or how enterprise customers should handle contracts that require unmarked deliverables. Until those answers arrive, service providers are left assuming that anything touched by a new Claude model is permanently attributable.
The rollout also landed in the same week Google moved in the opposite direction, giving Gemini users a switch to hide the visible watermark on generated media while keeping invisible signals in place. Two of the largest model providers now agree on machine-readable provenance and disagree about what the humans should see. Regulators wrote the rules for machines. The audience for watermarks, for now, is software.