They are introducing an invisible watermark to the texts written by Claude AI, here is how it will all work

This move represents a significant step towards greater transparency in the field of artificial intelligence and is a direct response to increasingly stringent regulatory requirements placed on technology companies.

How the invisible water žig works

Large language models, the company explained, generate text by choosing one word at a time. from the list of potential, appropriate words. It is common for them to choose words at random, as long as they make sense in the given context. When watermarking is enabled, Claude will use a special key; in order to decide which word to say; to choose, instead of using an arbitrary random number generator to choose the next word.

In his example, Anthropic used the digits of pi as the keyč. Let’s say that the keyč starts with digit two from the sequence 3.1415926535. The next generated word is it will be sixth on the election list, then fifth, third and fifth again. This method is an adaptation of the SynthID-Text approach developed by Google DeepMind, which the team described in a paper published in the journal Nature. Anthropic claims that the water feature does not affect the quality of Claude’s results or slow it down, and it will not require additional tokens or make content generation more expensive.

Response to European Union regulations

The implementation of this technology is a direct consequence of the European Act on Artificial Intelligence, namely Article 50, which provides transparency for content generated by artificial intelligence. The aim of the regulation is to suppress the spread of so-called “deepfake” content and misinformation by making it clear to users when the content was not created by a human. Failure to comply with the Artificial Intelligence Act may result in significant financial penalties.

Anthropic is not the only company adapting to the new rules. Other tech giants, including Google, Meta, Microsoft and OpenAI, have also signed up to the EU Code of Practice on the Transparency of AI-Generated Content and are implementing similar technologies to flag and trace the origin of content. This industry-wide trend signals a growing need to establish clear provenance of digital content in an era of increasingly sophisticated generative artificial intelligence.

Limitations and divided reactions

Anthropic acknowledges that their method of tagging text has limitations. Watermark can indicate the probability that Claude was involved in the creation or editing of the text, but it cannot definitively prove it. For example, if you ask Claude to just edit or translate something you wrote yourself, the text will also be highlighted. As they explain, I can only say that Claude was probably in contact with the text at some point.

In very short texts or in content that has been significantly edited after generation, watermarks may not be detectable. On the other hand, lightly editing the text generated by Claude probably won’t remove the tag. If you want to be sure, you will have to completely rewrite it. The announcement caused different reactions. Some users have expressed concern that their own work could be labeled as AI-generated if they use Claude for editing or translation. Still, others see this as a positive step toward transparency and a way to prevent AI models from being trained on their own results, a phenomenon sometimes called “model collapse.”

Impact on program code and images

Questions have also been raised about how watermarks will affect program code, given that it may not be subject to copyright without significant human input. If it could be proven that the entire codebase was generated by artificial intelligence, someone could copy and build upon it. Anthropic said the code has “generally fewer watermarks than some other forms of text” because it usually requires exact output. If there is no choice when generating the text, the watermark cannot be applied.

Anthropic will also tag the images by adding a cryptographically signed note to the metadata stating that they were generated by Claude. Images and other types of files will be marked according to the open standard C2PA (Coalition for Content Provenance and Authenticity), a system used by both Google and Adobe. The changes will affect all Claude products that use models released after August 2nd, and the company will add the tagging option to older models in the coming months.