Anthropic has shared new details on how the invisible watermark system that will be added to the content created by Claude will work. Developed to comply with the transparency requirements under the European Union’s Artificial Intelligence Act, the system aims to allow the AI-generated texts to be detected later.

However, this is not a classic watermark. No hidden characters or random marks noticeable to the user will be added to the texts created by Claude.

How does Claude’s invisible watermark work?

According to Anthropic’s explanation, the system works directly during the text generation process. When Claude chooses the next word, he selects from a large number of possible options. The new system creates a statistical pattern within the text by slightly altering the likelihood of certain options being chosen.

The aim is for these changes not to affect the meaning, quality, or readability of the text. Users will also not be able to directly see the difference between a regular Claude translation and text with a watermark.

Because the watermark is embedded in the text itself, it can be preserved after copy-paste processes and may be resistant to some modifications.

Anthropic also emphasizes that the watermark does not carry any random personal information. It will not be possible to link the content to a specific user, organization, or Claude conversation through the system.

Detection will be more difficult in short texts

One of the system’s significant limitations is the length of the text. Anthropic states that it is more difficult to reliably detect the watermark in short texts.

This is because the system relies not on a single word or character, but on a statistical pattern that emerges throughout the text. As the text gets longer, the system obtains more data to determine that it was created by Claude.

It is also stated that the watermarking process will not require extra tokens and will not create an additional cost for users.

Anthropic’s approach could be a valuable step in identifying the source of AI-generated content. However, the company also acknowledges that the system is not perfect. Detecting the watermark may be difficult in content that has been heavily rewritten or modified using different procedures.

So, do you think that detecting AI-generated text in this state is a realistic approach? You can share your opinions in the comments.