Anthropic adds invisible watermarks to AI-generated text
Anthropic will add invisible watermarks to Claude-generated text, helping detection systems more easily identify AI-created content.
AI-generated text is becoming harder to distinguish from human writing, and Anthropic is taking a new approach to make its origins easier to trace. Anthropic, the company behind Claude, has announced that it will add invisible watermarks to text generated by its AI models.
The technology is designed to help detection systems identify AI-generated or AI-edited content even when it looks like ordinary text to readers. This comes as AI companies face growing pressure from regulators, publishers and platforms to make synthetic content more transparent.
Why Anthropic is adding watermarks
Anthropic's move is linked to new European requirements around AI transparency. The EU AI Act's Transparency Code took effect on 2 August 2026 and requires AI-generated or AI-edited content to be marked in ways that can be recognised by other systems.
Anthropic says models released after 2 August will automatically include watermarking for computer-generated text and files. The company also plans to extend the technology to older models.
For files, Anthropic is using C2PA, an open technical standard that attaches content credentials to digital media. These credentials can provide information about where a file came from and whether it has been created or modified using AI.
How the invisible watermark works
Unlike a visible label, the watermark will not appear as a warning or disclaimer within the text. Instead, it is embedded into the generated output. Anthropic says the watermark can remain with text when it is copied and pasted elsewhere and may survive some forms of editing.
The company is applying the technology at the model level, rather than limiting it to a particular application.
That means the approach is expected to cover several Claude products, including the Claude platform API, Claude, Claude Code, Claude Cowork and Claude Tag. The idea is to make AI provenance part of the generation process itself.
What it means for users and publishers
For most users, the immediate experience may not change. Claude's responses can still appear like regular text on screen. The difference comes when the content is analysed by a compatible detection system.
Such systems could potentially identify whether the text originated from an AI model, giving publishers, businesses, schools and online platforms another way to assess content provenance. This could be useful as AI-generated writing becomes more common in news, marketing, education and business communication.
However, invisible watermarks are unlikely to be foolproof. This raises questions about how much editing would be required to weaken or remove Anthropic's watermark. Heavy rewriting or paraphrasing could make detection more difficult.
A wider push for AI transparency
Anthropic is part of a broader industry movement towards identifying AI-generated content. Companies including Google, Meta, Microsoft, OpenAI, Synthesia and Black Forest Labs have also committed to the EU's transparency framework.
The push reflects growing concerns about misinformation, automated publishing, unclear authorship and the rapid spread of synthetic media. Watermarking does not solve these problems on its own. It works best when combined with other measures such as clear disclosure, content credentials and reliable detection systems.
What happens next
Anthropic's move shows how regulation is beginning to influence the design of AI products. Instead of relying entirely on users to disclose when they have used AI, companies are increasingly building signals of origin directly into generated content.
The effectiveness of Anthropic's approach will depend on how reliably the watermark survives editing and whether other platforms adopt compatible detection technology. As AI-generated content becomes a normal part of digital communication, signs of its origin may increasingly travel with the content itself, even when those signs remain invisible to the person reading it.