I’ve seen a steady stream of articles in the past few days about the new text watermarking Anthropic announced for Claude. I got completely nerd-sniped by the concept of a watermark hidden in text, so I looked into it. It’s actually an interesting technical problem.

How it works

At every token in a generated LLM output, there’s a list of probabilities, and for the places where there are a bunch of equally legitimate choices, the model biases slightly in a deterministic direction. Over a lot of token choices, a 55% vs. 50% probability bias can be statistically significant, but hard for a human to notice. Pretty cool, but obviously brittle. I’m guessing you’d need a long enough stretch of consecutive Claude-generated text to clear that significance bar, so it likely wouldn’t flag workflows like cleaning up typos and light editing with AI — though a heavily AI-rewritten paragraph inside an otherwise human piece might still trigger it on its own. The validation API isn’t available yet that I could see, but it’ll be fun to play with some experiments once it ships.

Testing it on this post

Just for fun, I ran the two paragraphs above through GPTZero before publishing:

GPTZero result: “Mixed” and “AI Polished” — highly confident this text was human written and polished with AI

“Highly confident this text was human written and polished with AI.” That’s legit. I do lean on Claude for editing help on posts like this one. It’s also a decent illustration of where the detection landscape already is, watermark or no watermark: tools like GPTZero are already making this exact “human-written, AI-polished” distinction today, false positives and all.

The internet might be overreacting

I was surprised by the tone of a lot of these articles though. From what I can tell, Anthropic is simply trying to comply with a new EU law that requires this, so the negativity seems excessive. I’d guess it’s mostly people worried that they might get exposed for AI writing they’ve been passing off as their own — existing detectors like GPTZero have been catching some of that, but maybe folks who have elaborate systems to hide the AI use are angry that this closes the gap further. Or maybe there’s just a lot of Anthropic haters out there. Not saying these AI companies are always in the right, but this one seems like something they’ll all be forced down eventually.