Anthropic will begin applying invisible watermarks to text generated by its Claude models, the company confirmed this week, becoming the latest AI provider to add machine-readable provenance signals in response to new European transparency rules.
The change, detailed in an updated support page, means that content produced by Claude will carry a hidden signature designed to survive being copied and pasted elsewhere. It is a direct response to the European Union's AI Act Transparency Code, which took effect on August 2 and requires AI companies to mark generated or edited content in a way that other systems can detect.
How the watermark works
Anthropic said all models released after August 2 will automatically include technology to watermark both computer-generated text and files. For files, the company is adopting the C2PA open standard, a content credentials framework already used by some camera makers and image-editing tools to record how a piece of media was created.
The text watermark is embedded at the model level rather than applied as an afterthought. According to Anthropic's support documentation, "because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing. Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from."
The company added that it plans to extend watermarking support to older models as well. The feature will apply across Anthropic's product line, including the Claude platform API, the Claude chat app, Claude Code, Claude Cowork, and Claude Tag.
Exactly how much editing a user would need to do to strip the mark is not entirely clear, and Anthropic has not publicly detailed the technique. Text watermarking is technically difficult because small changes to wording can break fragile statistical patterns, a long-standing challenge for researchers in the field.
No opt-out for users
Notably, Claude users will not be able to disable the watermark. The setting is enforced by the model itself rather than toggled per request, which means anyone generating text through a Claude surface will have it marked whether they want it or not.
That stance is likely to draw pushback from some professional users. Writers, developers, and businesses that use Claude to draft content have grown accustomed to clean, unmarked output, and some may be uncomfortable with an invisible signature following their text around the internet. Anthropic appears to have concluded that regulatory compliance outweighs that preference.
Part of a broader industry scramble
Anthropic is not acting alone. The EU's transparency requirements have set off a race among AI and content platforms to add provenance signals before regulators come knocking.
Just last week, AI music platform Suno said it would begin marking tracks created on its service after a wave of legal challenges over AI-generated audio. Last month, newsletter platform Substack partnered with detection firm Pangram to flag AI-generated content, with Substack CEO Chris Best publicly calling out "Claudefishing," his term for people using AI to flood platforms with low-quality material.
The pressure extends well beyond Anthropic's direct competitors. Black Forest Labs, Google, Meta, Microsoft, OpenAI, and Synthesia have all committed to adhering to the EU's transparency code, according to reporting by TechCrunch, signaling that watermarking is fast becoming a baseline industry expectation rather than a differentiator.
Why provenance is suddenly urgent
The push for watermarks reflects two converging anxieties. The first is regulatory: the EU AI Act is now in force, and companies that fail to comply face the prospect of significant penalties. Getting provenance infrastructure in place early is cheaper than retrofitting it under enforcement pressure.
The second is reputational. As AI-generated text, images, audio, and video have flooded social platforms, publishers and platforms face growing backlash over "AI slop," low-quality synthetic content that crowds out human work and erodes trust. Invisible watermarks offer a technical answer to a question that moderation alone cannot solve: was this written by a machine?
For Anthropic, which has built its brand around AI safety, adopting watermarks is also a credibility move. A company that markets itself as the responsible alternative has strong incentives to lead on transparency rather than be seen dragging its feet.
The harder questions remain unresolved. Watermarks only help if detection tools actually look for them, and if the marks prove robust enough to survive the editing, reformatting, and laundering that happen as content moves across the open web. Anthropic's bet is that embedding the signal at the model level is the most durable approach. Whether it holds up in practice will be tested quickly, and publicly, in the months ahead.
For ongoing coverage of AI policy and the tools reshaping the industry, bookmark AI Buzz Wire's breaking AI news coverage.
