compartmentalization comments on What can go wrong with the following protocol for AI containment?

[compartmentalization]

If you use the entities’ preferences to decide what’s ethical, then everything is (or can be), because you can just adjust their preferences accordingly.