After Claude, ChatGPT’s hidden watermark is here too. Simply put, a watermark means AI generates content following a set of rules only it knows, slightly favoring certain words and formats. This ensures that detectors can later identify the content as AI-generated. What’s interesting, though, is that their announcement is practically a textbook on how to remove the watermark. For example, the shorter the text generated in a conversation, the harder it is to detect. Using 200 tokens (about 150 English words or roughly 200-300 Chinese characters, equivalent to a short paragraph), the detector has about an 80% chance of identifying the watermark. But if the generated text is shorter than that, the detection probability drops sharply. Additionally, OpenAI mentioned that if users manually rewrite the generated content, it significantly weakens the detector’s ability to identify AI-generated text. For instance, replacing just 10% of the words drops the detection rate from 92% to 66%. Replacing 25%? It plummets to just 17%. The good news is, for now, only ChatGPT and Codex users in the EU have their generated text subtly watermarked. But since this watermark is easy to remove, OpenAI isn’t making the detector publicly available to everyone just yet.