Students and even some professionals trying to pass off AI-generated work as their own may soon have a harder time getting away with it.
Anthropic announced this week that it will start adding machine-readable watermarks to content generated by its chatbot Claude, including invisible marks embedded directly into AI-generated text.
With this announcement, Anthropic joins OpenAI and Google in outlining how it plans comply with transparency requirements under the European Union’s Artificial Intelligence Act.
According to an Anthropic support page, new Claude models launched in the EU on or after August 2 will support the marking system from launch. Generated text will contain embedded watermarks, while supported files will carry digitally signed provenance metadata.
But the system won’t be limited to Europe. Anthropic says the marks will apply to supported models across “Claude Platform (API), Claude, Claude Code, Claude Cowork, and Claude Tag, and wherever Claude is offered, worldwide.”
The company says it is also working to add marking to existing models and plans to provide tools that will allow users and third parties to detect Claude’s marks.
When it comes to text, Claude will hide a machine-detectable pattern directly in the words it generates.
“When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself. You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response,” the company wrote.
Because the watermark is embedded in the text itself, it travels with the writing when it is copied and pasted elsewhere and may even survive some editing.
Still, there are some major caveats.
Anthropic warns that finding a watermark only indicates that the content may have been processed by Claude. Because people also use Claude to proofread, translate, summarize, or otherwise edit their own writing, text that originated somewhere else could still carry a Claude watermark. Marked text could also have been modified or combined with other material after Claude processed it.
The reverse is also true. The lack of a detectable watermark doesn’t necessarily mean something wasn’t generated by AI. Claude-generated text may lose the signal if it is heavily edited, paraphrased, translated, or mixed with other writing. Short passages may also not contain enough text for a reliable signal.
Anthropic isn’t the first major AI company to experiment with this kind of text watermarking. Google already uses its SynthID technology to embed invisible watermarks in AI-generated text.
Meanwhile, OpenAI currently uses transparency tools including SynthID for images and audio, but hasn’t pubilcally announced a detection system for text.
Beyond text, Anthropic will also start attaching signed provenance metadata to supported files such as .svg, .png, and .jpg. The metadata follows the Coalition for Content Provenance and Authenticity (C2PA) open standard and can signal that a file was processed by Claude and whether it has been tampered with.