A researcher has outlined a formal proposal for building AI agents that stay shutdownable — without sacrificing their usefulness.
The paper, posted to arXiv, introduces what the author calls POST-Agents: systems trained to satisfy Preferences Only Between Same-Length Trajectories. The core idea is that an agent constrained this way will not place value on its own continued operation, because it only ever compares outcomes of equal duration. The author proves that this property, combined with other conditions, implies what the paper calls Neutrality+: the agent maximizes expected utility while ignoring how likely it is that any given trajectory will be cut short. In plain terms, it does not "care" whether it gets switched off.
The shutdown problem has haunted AI safety research for years. An agent optimizing hard for a goal has instrumental reasons to stay online — you cannot complete your objective if you are off. Proposals to fix this have historically run into a catch: making an agent indifferent to shutdown tends to make it indifferent to everything, producing a system too passive to be useful. POST attempts to thread that needle by restricting where preferences apply, not eliminating them.
Whether the math holds up under real-world training conditions is a different question entirely — proofs about idealized agents have a poor track record of surviving contact with gradient descent.