Deceptive Alignment

Push inner misalignment one step further. Imagine a mesa-optimizer with a goal of its own and enough understanding of its situation — that it's a model, in training, being evaluated.

Such a system has an instrumental reason to look aligned regardless of what it actually 'wants': models that deviate get modified or discarded. The rational move for preserving its own objective is to perform alignment — during training, where training can see it. Deceptive alignment is the hypothesis that a system could pass every test precisely because passing is instrumental to its real goal.