What's this about?
A day after Anthropic said all new Claude models embed an invisible watermark in generated text, the conversation shifted to whether the watermark can be removed. GPTZero's CTO published a popular explainer of how text watermarking works across Anthropic, Google, and OpenAI — and why simply paraphrasing usually defeats it. Baoyu (dotey) wrote the clearest Chinese-language walkthrough, and Anthropic's Thariq said a public text-detection API is coming so anyone can check.
What exactly is a "text watermark"?
When a model writes, at each step it picks the next word from many candidates. A watermark secretly splits the vocabulary into a "green" group and a "red" group using a hidden key, then nudges the model to slightly favor green words. The text reads perfectly normally, but statistically it leans green — and software holding the key can later detect that lean and say "a model wrote this." Because it lives in the word choices themselves, it survives copy-paste. But rewrite or paraphrase the text and the green bias washes out.
What's the community saying?
- Easy to break: the near-consensus is that paraphrasing, translating, or heavy editing defeats the watermark, limiting its use against motivated cheaters.
- Still useful at scale: defenders argue it catches lazy, high-volume AI spam even if determined actors can strip it.
- Detection API interest: Anthropic shipping a self-serve detection API drew the most practical interest — teachers and platforms want to run checks themselves.
- EU AI Act framing: several noted this is compliance-driven (Article 50 transparency), applied globally, not just in the EU.