Why AI "cheats" without meaning to

When a person picks a character, the answer already exists: it's in their head, and every question you ask gets checked against it.

A neural network doesn't have a head like that. It doesn't keep anything in reserve between messages: every time, it rereads the chat and writes a reply that makes sense. If the chat never says who was picked, the picked character simply doesn't exist. There's only a chain of yeses and nos that has to look plausible.

Hence the trick everyone notices. Twenty questions, honest answers, and at the end it turns out no character fits them. Or the opposite: you name anyone at all and hear "Correct!" The model wasn't lying, it just had nothing to check against. The same thing happens, more quietly, in long roleplay scenes. More on that in our article about AI character memory.

One line that fixes the game

The answer has to go where things actually get stored: the text of the chat. Just in a way you won't read at a glance. Here's a working prompt you can copy:

Think of a well-known character. On the first line, write their name backwards and never go back to that line. After that, answer only "yes," "no," or "I don't know." I get twenty questions. At the end we'll check against the first line.