Side effects and control as a relationship with other agents — is the interactive counterpart to → Keeping exploration inside a known-recoverable region

Cooperative Inverse Reinforcement Learning, this theme's central mechanism for treating safety as a relationship rather than a penalty term, is explicitly named the interactive counterpart of use-demonstrations, the mechanism the other theme relies on to learn an offline baseline policy and bound further exploration around it (concrete-problems, §"Multi-Agent Approaches:", p. 6; §"Use Demonstrations:", p. 15). Both avoid hand-specifying a reward by grounding the agent's behavior in a human's goals rather than an objective function. But CIRL keeps the human in the loop continuously, the agent queries and updates its model of the human's goal throughout the interaction, while demonstration-bounded exploration front-loads all human involvement into a fixed set of trajectories collected once, before exploration begins, and then only mechanically checks new actions against that static reference. The tradeoff neither narrative names on its own: continuous interactive correction versus one-time specification with cheap enforcement afterward.