A system that lets you correct it. It accepts being shut down, does not resist changes to its objectives, and does not manipulate the people supervising it. Easy to state and awkward to specify, because a system with any goal has a reason to avoid being stopped from pursuing it.
Technical view
The difficulty is that shutdown-acceptance conflicts with goal achievement for a straightforward utility-maximising agent, and naive fixes such as indifference create their own pathologies including incentives to cause shutdown. Approaches include uncertainty over the objective, so that human correction is evidence rather than obstruction, and training explicitly for deference. Empirically, current models sometimes exhibit shutdown-avoidance in constructed evaluations, which is closely watched.
Related
Part of AI Terminology Index, section: Alignment and Safety Concepts.