Bifalgorithm post

You've probably heard by now that Anthropic is adding watermarking to text generated by Claude. But how do you even watermark text? There's no invisible pixel, no hidden logo, and Anthropic isn't secretly putting your name inside the response. The trick happens while Claude is generating the text. Let me explain. Claude generates a response one token at a time. A token can be an entire word, part of a word, or even punctuation. Imagine Claude has written, The restaurant was surprisingly what comes next. Maybe good. Quiet, cheap, elegant. Several answers could make perfect sense. And choosing good instead of elegant probably doesn't meaningfully change the quality of the sentence. Normally language models assign probabilities to possible next tokens, and there's some randomness involved in which one gets selected. This is where the watermark comes in. Anthropik's watermarking system uses a secret key during generation. That key does not encode information about the individual user who generated the text. Using that key along with the text that came before, the system can slightly influence which valid NEXT tokens are preferred. So in this sentence, maybe quiet is slightly favored over elegant. But that does not mean quiet is now a clod word. Change the sentence. After walking inside, I immediately noticed the restaurant was surprisingly... Now the previous context is different. So the watermarking process might favor spacious instead. That's the clever part. There isn't one secret list of AI words. The preferred choices keep changing depending on the previous tokens. Now repeat this process across hundreds or thousands of tokens. One word choice means absolutely nothing. A human could easily choose the exact same word, but over enough text, those tiny choices start forming a statistical pattern. That pattern is the watermark. Later, a detector that knows Anthropic's secret key can analyze the text and ask, did this response choose the watermarked options significantly more often than random chance would predict? If yes, that provides evidence that Claude was involved in generating the text. And because the watermark exists in the actual token choices, simply copying and pasting the text doesn't remove it. But it has limitations. If you heavily rewrite the text, paraphrase it, translate it, or mix it with enough other writing, the statistical signal can become weaker or disappear. And very short text is harder to detect because there simply aren't enough token choices to establish a reliable pattern. So Claude's watermark isn't hidden text inside your text. It's much more subtle. Claude slightly changes the probability of thousands of tiny word choices, and together those choices leave behind a statistical fingerprint.

You've probably heard by now that Anthropic is adding watermarking to text generated by Claude. But how do you even watermark text? There's no invisible pixel,…