How does this text 'watermarking' generally work?

Text watermarking works by subtly biasing the probabilities the model assigns to different word choices during generation, following a specific statistical pattern that detection software can later recognize, even though the text reads completely naturally to a human. Because the bias sits in patterns of word choice rather than any visible symbol, it doesn't change how the writing looks or feels, but leaves a signature detectable by a tool designed to check for it.

It does not involve printing a visible stamp on the text, since the point is that it stays invisible to readers while still detectable by software. It also doesn't embed personal information like your name; the pattern is a statistical fingerprint, not stored personal data.

Because the watermark relies on a fairly specific pattern, heavy paraphrasing, translation, or substantial editing of AI text can weaken or erase the signal, one reason it is a useful but imperfect tool rather than a foolproof way to catch AI writing.