How AI text watermarking works

Most of the useful reaction was not about the math. It was about where this actually works in the market we have. People largely agreed this is weak as a general-purpose detector because anyone motivated enough can route through a different provider, use an open-weight model, or simply paraphrase the result. That makes it a poor answer to plagiarism, disclosure, or broad provenance. Where it does look real is inside the vendors’ own workflows. A provider can use it to estimate how much of the web contains its own outputs, keep synthetic text out of future training sets, and potentially sell or gate access to detection APIs. Several commenters also pointed out the structural consequence: because detection requires the model and the secret key, closed-model vendors end up with a monopoly on detecting their own output. A second theme was quality. Many people used the article as a hook to complain about Claude’s recent writing style, describing it as patronizing, over-metaphorical, and sometimes incoherent. But the stronger point was that watermarking is not a satisfying explanation for that. Anthropic says only newer models support it, Gemini has its own watermarking approach with a different voice, and the mechanism itself can be subtle enough that it need not obviously distort prose. The more grounded conclusion was that watermarking may impose some tradeoff at the margins, especially in constrained domains like code or short factual snippets, but it does not by itself explain every bad model-writing habit users are seeing. The hardest unresolved piece was code. In prose, there are usually many acceptable token choices, so a gentle bias can hide in the distribution. In code, commands, syntax, identifiers, and short snippets often leave much less room to maneuver. That led many people to doubt whether watermarking code is robust enough to be useful, except perhaps in comments or longer natural-language sections. The result is a technology that feels less like a universal AI lie detector and more like a private provenance feature for big labs, with some regulatory cover and some obvious business upside.