Every piece of text and every image generated by Anthropic’s Claude will soon carry a hidden signature. The company announced Monday that it is adding invisible watermarks to all model outputs, along with C2PA metadata, in what it described as a global, mandatory change that users cannot turn off. The move makes Anthropic the first major AI company to apply content provenance technology to every output of its flagship products.
The watermarking requirement traces to a specific source: the implementing rules of the European Union’s AI Act, the bloc’s sweeping regulation of artificial intelligence. Anthropic said the change applies worldwide rather than only in the EU, a choice that converts a regulatory compliance deadline into a product feature. Fortune described the decision as one of the most aggressive moves yet on content provenance, and the industry is watching how other labs respond.
The technology works in layers. An invisible watermark embeds a machine-readable signal in the output itself, designed to survive editing, cropping and reformatting. The C2PA metadata, built on a standard developed by a coalition including Adobe and Microsoft, records where and how the content was made, creating a chain of custody that travels with the file. Together the two layers give verifiers two ways to check whether a given text or image came from Claude — and, in principle, no way for the model’s maker to deny it.
The policy changes the industry’s direction of travel. For two years the dominant approach to AI content has been detection: building tools that spot machine-generated material after it appears. Anthropic’s approach is different. Instead of waiting to detect AI content, the company is marking the content itself at birth, so that provenance is a property of the thing produced rather than a judgment made about it later.
The economics of the decision are not obvious. Watermarking is a cost center: it adds engineering complexity, risks degrading output quality, and gives competitors a compatibility problem to solve. Anthropic is absorbing those costs voluntarily, worldwide, ahead of any requirement to do so outside Europe. The company’s stated rationale is trust — that provenance will become a selling point as AI-generated content floods the web and customers demand to know what is real.
There is also a defensive logic. The EU’s AI Act carries penalties for noncompliance, and the companies building frontier models face the deepest scrutiny. By implementing the rules early and globally, Anthropic gets to shape how the standard is interpreted, sets the terms of the debate, and presents itself to regulators as the cooperative player in a field that has often treated regulation as a threat. The strategy worked for the company before, and the watermarks extend it into the compliance era.
The harder questions are practical. Text watermarks are harder to make robust than image watermarks, because text can be paraphrased, translated and rewritten while keeping its meaning. Researchers have shown that watermarking schemes can be evaded, and that the evasions often come from the same model makers — a dynamic that has made some in the industry skeptical of the entire approach. Anthropic says its system is designed to be resilient, and the company is expected to publish technical details so outsiders can test the claims.
For customers, the change is subtle and mostly invisible: watermarks are designed to leave output quality and utility intact, and most users will never notice them. The visible effect is institutional. Enterprises that have hesitated to adopt AI tools because they cannot prove which content is machine-made get a mechanism for exactly that proof, and publishers negotiating over AI-generated material gain a technical record to point to. Anthropic is betting that this trust dividend outweighs the cost of the technology, and that the market will pay for provenance once it exists at scale.
For the rest of the industry, the move sets a benchmark. OpenAI and Google have each experimented with watermarking in their own products, but neither has announced the kind of blanket, global, no-opt-out policy Anthropic just described. If the industry converges on Anthropic’s standard, provenance becomes table stakes rather than a differentiator, and the cost of the technology moves from voluntary to unavoidable.
The long-term stakes are about what the web becomes. If every major model labels its output, the flood of synthetic content becomes traceable, and platforms gain a tool they have been asking for: a way to sort machine-made material from human-made material at scale. Anthropic has chosen to build that tool first, for everyone, whether they asked for it or not. The company’s bet is that provenance is the next trust layer of the internet, and that the lab that builds it first gets to define it.
