Anthropic releases more info about Claude AI watermarking amid user confusion

Days after Anthropic announced that text content generated with its Claude AI would include an invisible watermark, the business has released more information concerning how the marking technology works.

Anthropic refuted users’ theories that machine-readable characters might be inserted into the AI-generated text, confirming that nothing would be added to the text and there were no hidden characters.

Instead, the watermarking system relies on the AI leaving a pattern in Claude’s responses. It is detectable to any party that has a key that encodes it.

According to Anthropic, claude’s text watermark was a version of the SynthID-Text approach published by Google DeepMind in a Nature paper two years ago. The company repeatedly stressed that the AI watermarking process would not impact the quality of Claude’s content, level of creativity, or text readability.

Watermarking will also not affect the speed or rate of using the AI models.

Regarding the visibility of the marking, the Claude AI watermark has a chance to evolve and become detectable the more the AI is used. In other words, the more Claude writes, the more “space” there is for a watermark, per the company.

Naturally, longer passages are more likely to have a watermark when compared to a shorter passage that Claude has processed. According to the company, code “generally” had less watermarking than some other forms of text, but that translations would carry a Claude watermark because every word was chosen by the AI.

Anthropic stated in its blog post: “A watermark can only determine that Claude was likely involved with the content at some point. It cannot distinguish “Claude wrote this” from “Claude heavily edited this.”

Anthropic again listed the limitations of watermarking text, even as users expressed confusion and concern over the possibility of the Claude AI watermark leading to false positive or false negative results that could hurt writers’ careers.

In response to questions about whether editing the text would be enough to evade a watermark, Anthropic responded that light editing “probably” won’t remove the watermark completely, but replacing every word of the text could do so.

The watermark cannot be used to trace back to a specific individual, organisation, or chat session where the text was generated.

Over the coming months, Anthropic is due to add watermarking for its older models as well.

The Google and Amazon-backed company reiterated that it was implementing the watermark to comply with the EU AI Act, stressing that a number of other “major” AI model providers signed the EU Code of Practice on Transparency of AI-Generated Content in July this year.