Anthropic is adding imperceptible, model-level watermarks to Claude's AI

AI-generated text, images, and video have all reached that level where it's becoming increasingly difficult to tell the difference between what's human-generated versus stuff spat out by hardware sitting in a data center somewhere. This is why the European Union's Artificial Intelligence Act (AIA) requires AI companies to implement identification measures to ensure there's a simple check, while offering a Code of Practice.

And on that note, Anthropic has posted a new article on its Claude support portal titled: How Claude marks AI-generated content. Here it details how Claude will begin to embed watermarks in AI-generated text, which the company describes as imperceptible and injected directly into the text itself. It will reportedly persist even with some editing, meaning that for those who do some light editing and reword Claude-generated text, the result will still include the watermark identifying it as AI-generated content.

"Because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing," Anthropic confirms. "Watermarking will be applied at the model level, which means it will be present no matter which Claude product or surface the text comes from."

Anthropic notes that it's also working to enable Claude users and third parties to detect the embedded watermarks, with details on how this will work still to come. Anthropic notes that a positive detection means that it will mark text as being "processed by Claude," but it won't be "fully conclusive." One reason is that someone using Claude to proofread or make changes to a document doesn't necessarily mean Claude is the original author.

TweakBot answers common questions about this news using TweakTown's own coverage from this page and related content from our archive. All answers are generated from TweakTown content and not outside sources. Tap a question to reveal the answer, or type your own below related to this content.

Have a question about this content not listed here? Ask below and TweakBot will answer it.

Interestingly, when it comes to AI-generated images, Anthropic's Artificial Intelligence Act (AIA) compliance doesn't actually embed a watermark in the .svg, .png, or .jpg images you see, but rather via attached signed provenance metadata. Basically, the metadata in the file will include the confirmation that it was processed by Claude. When it comes to copying an AI-generated image via a screenshot and then saving it manually, this digital certificate of sorts is lost, so it's not as foolproof as embedding the watermark in generated text at the model level.