August 13, 2026
Marked for Drama
How AI text watermarking works
AI’s secret text tag is here, and commenters are already calling it obvious, creepy, and a cash grab
TLDR: AI companies are hiding a detectable fingerprint in word choices, letting them later guess whether text was machine-written even after copy-paste. Commenters are split between "this is obvious," "this sounds creepy," and "here comes a paid detection racket," which matters because it could shape who gets to label content as fake or real.
The big reveal in How AI text watermarking works is almost annoyingly simple: the "mark" isn’t hidden in invisible code or mystery pixels, because plain text doesn’t have those. Instead, an AI leaves a faint fingerprint by slightly favoring some perfectly normal word choices over others. To regular readers, the sentence looks the same. To the company holding the secret key, it can allegedly be counted and flagged later. That’s the science. The comments? Absolute chaos.
Some readers reacted with pure disbelief that this was treated like a grand secret at all. One commenter basically said a math nerd could invent this over coffee, while another flexed that they figured out the idea in five minutes. Ouch. Others immediately jumped past the "how" and went straight to the money and power drama: if companies can detect whether Claude or Gemini wrote something, does that become a paid verification service? A neat safety tool today, a subscription portal tomorrow? That suspicion lit up fast.
Then came the political paranoia and sarcasm. One person framed it as American AI companies watermarking text to satisfy regulators in Brussels, with only elite firms getting access to the magic label-maker. Another simply sneered, "AI writing at its finest," which is basically the comment-section version of throwing a tomato. The vibe overall: half the crowd thinks this is clever but obvious, the other half thinks it’s a future toll booth for truth online, and everyone agrees the real drama starts when edited text, copied text, and paid detection tools collide.
Key Points
- •The article explains that AI text watermarking hides a statistical signal in model word-selection patterns rather than in visible characters or metadata.
- •It states that Google has watermarked text from the Gemini app and web experience since 2024, while its API was a documented exception at the time of writing.
- •It states that as of August 2026, new Claude models were marking text at the model level, with earlier models being added afterward.
- •The described watermarking method uses a secret key to partition candidate next words into groups and slightly bias sampling toward one group.
- •Detection works by using the same key to reconstruct those groupings over a text and counting whether favored words occur more often than chance.