Claude will start watermarking everything it writes
When you purchase through links on our site, we may earn an affiliate commission. Here’s how it works.
To comply with the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content, Anthropic has announced that “New models will mark AI-generated content from day one”.
This is a remarkable step for Claude, because it not only applies to any images it generates, which are relatively easy to watermark, but also to any text it generates.
Anthropic says that Claude models will have an “imperceptible watermark” embedded directly into generated text at the model level. It says the mark should survive copying/pasting and minor edits.
Anthropic has not yet publicly explained the exact algorithm or released a detector, so we can only speculate for now about how it might be doing this and its effectiveness.
Personally, I think that Claude is about to start hemorrhaging customers, unless the marking is relatively easy to circumvent, or all the other major AI players immediately follow suit.
If every piece of text it produces will now be easily identified as AI, then Claude becomes useless as a tool to a lot of people who are currently using it to generate text and are not being entirely honest about where that text came from.
And then there’s the issue of using Claude to proofread your human-written text — will your text now be flagged as AI if you accept Claude’s editing advice?
Sign up for breaking news, reviews, opinion, top tech deals, and more.
Your initial reaction to that might be “Good! You should be forced to reveal when AI has written something, and it’s about time people started writing on their own again!”, and you’d be entirely justified in that opinion.
But while it remains the only one of the big three AIs that’s doing this, I think we’ll see a lot of people switch to either ChatGPT or Gemini, because they don’t want to be revealed as using AI in their work.
We don’t know exactly how Anthropic is doing its marking with text yet, but my best guess is statistical watermarking during token generation rather than hidden Unicode characters or metadata. Imagine that at every point Claude is choosing among several perfectly reasonable next words:
e.g. The movie was excellent / superb / terrific / impressive.
Normally it chooses according to the model's probability distribution. A watermarking system can secretly divide possible tokens into preferred and non-preferred groups using a key. Claude then gives a tiny statistical nudge toward the preferred group.
