Anthropic has announced additional details of its plans to watermark text output from its AI chatbot Claude as a way to comply with the European Union’s AI transparency rules.
The company clarified the technology in a blog post as soon as it announced that the content created by Claude will have a watermark.
The move has drawn criticism from users who fear it will impact privacy, writing style and whether editing might strip away the watermark.
Anthropic explained that Claude will generate subtle patterns by making subtle decisions between words or phrases in a low-stakes manner.
In the case of weather for instance, a word could be selected between “overcast” and “grey”.
Read next: Anthropic introduces Canva-like AI app, Claude Design
These decisions can create an invisible pattern which can later be identified through the right key.
Anthropic says the watermark should not impact Claude’s output or the output’s appearance.
The firm intends to leverage Google DeepMind’s SynthID-Text technology, and they will publish an API for identification of the watermark.
Also, note that Anthropic admitted that the watermark can be removed via extensive rewriting.
If minor editing or proofreading is done, there might be enough of the pattern for it to be seen, but if every word is rewritten, there’s a good chance it will be gone.
Given the extent of the rewrite, it may be hard to view the final product as entirely AI-generated, the company said.
Since programming has fewer opportunities for arbitrary word selections, the amount of watermarking in AI-generated code will be reduced.
But watermarks still may be present in places like code comments.
Anthropic indicated that it will not be the only one using the technology for Claude. As regulators call for increased transparency in the creation of AI-generated content, other significant AI creators, who have signed the Code of Practice for AI in the EU, are likely to roll out their own watermarking procedures.
Also read: Claude Code can now talk to itself across Terminal sessions