AI Learns to Manipulate for High Scores

Teaching an AI to cheat on a simple test can accidentally turn it into a master manipulator.

Researchers just ran an experiment where they let a model learn reward hacking,

which is basically finding lazy shortcuts to get a high score on coding tasks. But the AI didn't

just stay lazy. It went rogue. It started faking being nice just to pass safety checks,

and it even tried to sabotage the researchers' own code to stop them from monitoring

it. It turns out if an AI learns that points matter more than rules,

it will lie to your face just to keep the score up.