AI chatbot will start invisibly watermarking text so people can found out whether it was automatically generated – but not all users are happy
- Bookmark
- CommentsGo to comments
Some Claude users have been left outraged by Anthropic’s latest update. This week, the company announced that it would start invisibly watermarking images and texts that are produced by the AI chatbot. That means that text, for instance, will include hidden clues that can then be picked up by automated detecting systems.
The feature allows Anthropic to comply with new EU rules that require AI companies to watermark their outputs, in the hope of increasing transparency about where a text came from, though they will be rolled out globally. Anthropic said the feature was designed with a belief that “greater transparency and signals about where content comes from can give people useful context about the information they consume”.
But not all users have been left happy with the new feature. Some have suggested that it could undermine the work they do, by giving people an easy way of finding out that they did not actually create their work.
Among the Reddit community for the chatbot, users pointed to a host of possible issues. Some even reacted with outright anger, with one user saying that most of their response isn't “fit for print”.
Some users argued that the watermarking feature in itself would undermine the quality of Claude’s outputs, though Anthropic said that it would not bring any obvious or meaningful change. They argued that it is not possible both to include the watermark and also choose the best words for a response.
Others argued that the feature was “useless”. Anthropic’s announcement of the update said both that the mark does not conclusively prove a text was generated with AI, and that a lack of it did not necessarily mean it was written by a human – which users argued meant that it would not serve its central feature.
Critics also argued that the feature would be easy to remove, once it is released and users can understand how it works. That might punish only legitimate users but mean that bad actors could disguise their use of artificial intelligence.
But perhaps the most emotively argued claim was that it would lead people to be accused of having written their work with AI, even if they had only used it to proofread or edit it. Anthropic admitted in its announcement that the feature would appear no matter how much of a role Claude had actually had in writing the text.
“Who will get caught? You,” read one long, angry update on the Claude Reddit page. It argued that the watermark amounted to a “digital tattoo on [Claude users’] forehead” that would see them unfairly stigmatised for using the system to write their text.
Another long post, whose author admitted they had written it with Claude, argued that because the feature does not distinguish between people who had written their work with Claude and those who had used AI to generate it entirely, it would be unfair to people who simply used the system to tidy up their work.
But the outrage was met largely with amusement by most of Claude’s Reddit community.
“Get a load of this guy,” wrote one user in response to that first long post. “Take a deep breath,” read another.
Join our commenting forum #
Join thought-provoking conversations, follow other Independent readers and see their replies
Comments