Anthropic has begun to introduce a system that will allow identify by machines the texts generated by the new models of Claude. The company will incorporate an imperceptible watermark for the user that will be part of the text itself and can remain even after copying and pasting it elsewhere.
The measure responds to article 50 of the European Union Artificial Intelligence Law which from August 2, 2026 obliges generative system providers to make synthetic text, image, video or audio content be identified by machine-readable markings. Systems marketed before that date have a transitional period until December 2 to fulfill this obligation.
Anthropic will apply the new system to Claude models launched from August 2 and ensures that the marking will be active worldwidenot only within the European Union. The company also works to incorporate it into previous models during the transition period provided for by the regulations.
The mark will not be a visible notice or hidden text that can be located when opening a document. Anthropic explains that it will be introduced an imperceptible signal during text generation and at the model level, so that it will travel with the words when an answer is copied and pasted. The company has not yet published the technical details necessary to explain exactly how it will encode that signal and prepares tools so that third parties can detect it.
This method will also not turn any text by Claude into infallible proof that it was written by an artificial intelligence. Important modifications made later can weaken or eliminate the signal and fragments that are too short may not contain enough information to detect it reliably. An unmarked text also does not necessarily prove that it was written by a person.
We must not forget that, in addition to generating text, Claude can be used to translate, summarize or correct one originally written by a person and the result could equally incorporate the model’s signal. The mark would therefore demonstrate the intervention of Claude, not who originally wrote the content.
The files will follow a different system because they allow information about their origin to be incorporated without modifying their content. Claude will use in compatible formats, including images, signed metadata based on C2PAan open standard Coalition for Content Provenance and Authenticity that allows you to verifiably record which tool created or modified a file and part of its digital history. These data also have limitations because they can disappear when you convert the file, save it again, or take a screenshot.
Anthropic is not the only large company that is adopting these technologies to meet new European requirements. Google already uses SynthID to mark the texts generated by Gemini through small modifications in the probabilities with which the model selects the following tokens, creating a pattern that is invisible to the reader but that can be detected later. He also uses it in images, but in a different way. Instead of statistically modifying the token choice, integrates an imperceptible signal directly into the image which can continue to be detected even after certain modifications.
OpenAI currently uses C2PA and invisible SynthID flags on compatible images generated with ChatGPT, Codex and his APIwhile some audio content also incorporates SynthID. The company has confirmed that it intends extend these provenance signals to all modalities, including textto fulfill its commitments to the Code of Good Practices European, although it has not yet explained what technology it will use or when it will specifically reach the texts generated by ChatGPT.