AI control

A branch of AI safety research that does not align a model's goals but checks its behaviour — whether it solved exactly the task it was given and nothing more; monitor models that watch other models and isolating models in containers belong to it.