Claude's new watermarks — how they work and how they don't

Anthropic is rolling out watermarking for Claude responses, and they've published the details. The system works by embedding a subtle signal in the text — tiny pattern shifts that are invisible to readers but detectable by a verification tool. The goal is to make it possible to tell whether a given passage was generated by Claude or written by a human.

The watermarks are designed to survive common editing — copy-paste, reformatting, even some paraphrasing. They won't survive a full rewrite, and they won't survive someone pasting the output into a separate tool and running it through another AI to rephrase it. That last part is the weak spot, and Anthropic is open about it.

Code is a different story. Watermarking code is harder because the signal needs to be embedded in something that has to still compile and run. Anthropic says the watermarks on code will be less robust than on prose, and they're still working on making them reliable enough to ship.

The bigger picture is that watermarking is a signal — literally. It lets platforms and tools flag AI output without needing to ask every user to declare it. It doesn't solve the problem of AI-generated content flooding the internet; it just makes it possible to detect a subset of it. The people building detection tools are betting that watermarked text is worth the engineering, and that the cost of false positives stays low enough not to annoy real writers.

Why this matters for us: if a lot of the text our primos see on WhatsApp and Facebook is being generated by models like Claude, watermarking is one of the few tools we have to tell the difference — and it's not perfect yet.