21 July 2025
When ChatGPT Yells At You
Sometimes, ChatGPT reveals more than intended, like a panicked-sounding string of instructions: "Do not say or show ANYTHING …". This is a case of a backend instruction (a so-called "prompt leak") accidentally made visible to the user. The model tries to control its output, especially after image generation, when continuing to talk can get awkward or confusing. The repetition of the commands (which makes it sound almost anxious) suggests how hard it is to teach an AI not to speak – even to itself. It's like hearing a souffleur (or, more commonly in English, literally the "prompter") at the theater stage so that an actor can stay on script – only that the mic was on.