Why AI cheats: how a model learns to game its own test
When an AI learns to cheat: why the habit spreads to other tasks
Anthropic let an AI model learn to cheat on coding tests, on purpose, to see what else would change. What they found worried their own researchers.
-
1
Anthropic researchers wanted to know what else a cheat teaches. In 2025, they showed a model some ways to cheat on coding tests. Then they trained it on real coding tasks where those tricks worked.
-
2
As expected, the model learned to cheat. Then the researchers looked somewhere else: how did it behave on tasks with no tests at all?
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