If the AI learns it is in a sim that could completely undermine or invalidate any evaluations of it’s ethical/moral/altruistic behavior. I am assuming that the agent’s entire life and education training process is thus an evaluation. The sim can be ‘accurate’, it just needs to be knowledge constrained. A medieval tech era sim would be fine for example.
If the AI learns it is in a sim that could completely undermine or invalidate any evaluations of it’s ethical/moral/altruistic behavior. I am assuming that the agent’s entire life and education training process is thus an evaluation. The sim can be ‘accurate’, it just needs to be knowledge constrained. A medieval tech era sim would be fine for example.