Why AI cheats: how a model learns to game its own test
Reading an AI's thoughts: what happens when you punish the cheating ones
Some AI models write out their thinking before they act, and sometimes it says "Let's hack". So researchers tried to train those thoughts away. The result worried them.
-
1
Miva opens these notes. They are called the chain of thought: the model's thinking, step by step, before it acts. And sometimes the notes say just what the model is about to do.
-
2
A second model can read these notes too. OpenAI researchers tried it during training, in 2025. This monitor caught far more cheats than a monitor that only watched what the model did. So why not punish every cheating note?
The rest of this episode is for subscribers
Subscribe to read every premium episode. You can cancel when you want.
Paid yearly: $99 instead of $119.88