3.1415926535—this string of numbers now has a new purpose. Anthropic explains that when Claude generates text, the model has to pick one word from a list of "plausible candidates" every single time. In the past, this was decided by random numbers; now, with the watermarking feature turned on, Claude instead uses a "key" to determine which position on the candidate list to pick from. The example Anthropic gives is pi itself: starting the key from 2, the following words are picked in order from the 6th, 5th, 3rd, then 5th option on each candidate list. The process doesn't leave behind any extra hidden characters, and there's no visible difference to the eye—only someone holding that exact key can decode it.
This approach is Anthropic's refined version of the SynthID-Text method that Google DeepMind published in Nature. Anthropic emphasizes that the watermark doesn't slow down generation or require extra tokens, meaning it won't drive up usage costs—this is a direct response to the EU's new AI transparency rules, which require AI-generated content to be identifiable through markings. Anthropic plans to release an API afterward, letting people use the key to check whether a piece of text was ever "touched" by Claude.
What can be detected is only "contact," not "how much was written"
The limits of the watermark are stated plainly too: it can't tell whether a passage was written by Claude from scratch or just lightly polished by the user's request. Translated text gets marked the same way. If the editing is light or the text too short, the watermark may become too weak to detect—but conversely, light editing essentially can't wash the watermark away. Anthropic's straightforward advice: if you really want it gone, you'll need to rewrite the whole passage.
Code is the exception. Anthropic says watermarks in code tend to be weaker than in regular text, for a simple reason: programming languages often demand precise output with little room for "candidate words," leaving no space for choice—so the watermark simply can't take hold. This ties into a real-world concern: people worried that if it could be proven an entire piece of code was AI-generated, copyright questions might just lead people to copy and modify it freely.
Images follow a different logic: Anthropic adds a cryptographically signed marker in the metadata stating the image was produced by Claude. As for why this is rolling out to users worldwide all at once rather than launching in the EU first, Anthropic's reasoning is that it currently isn't confident it can implement the feature separately by region—so instead, models released after August 2 will apply the watermark universally, while older models will get it added gradually over the coming months.
