Myopia (AI Alignment): Agents That Place No Value on the Future
A myopic agent values only the current task or episode, assigning nothing to what follows. Shutdown resistance, long-horizon deception, and resource accumulation all disappear as consequences rather than patches, yielding corrigibility as a by-product. The key contrast with shutdown-seeking designs is indifference rather than desire — no gradient toward the off-switch to exploit.
In AI alignment, **myopia** is a design property: an agent that values only the current task or episode and places no value on the future beyond it. The term is a technical borrowing and has nothing to do with the eye condition — for that see Myopia: Why Evolution Didn't Lock In Perfect Eye Shape. ## What myopia buys A myopic agent is *indifferent* between "finish the task" and "finish the task and then be shut down", because it assigns no value to anything after the episode ends. Several dangerous incentives disappear at once as a consequence rather than as separate patches: - **No shutdown resistance.** Continuing to exist has no value, so nothing is spent protecting itself or its off-switch. - **No incentive to deceive over time.** Hiding behaviour pays off only through future consequences the agent doesn't value. - **No resource accumulation.** Power, money, and influence are useful for future goals; a myopic agent has none. This yields Corrigibility: Building an AI That Doesn't Resist Being Corrected as a by-product rather than as an anti-natural property bolted on against optimisation pressure. ## Indifference, not desire The distinction that matters is between an agent that *doesn't mind* shutting down and one that *wants* to. Myopia produces the first; Shutdown-Seeking AI: Goldstein and Robinson's Beneficial Goal Misalignment produces the second. Wanting shutdown reintroduces a pull toward the off-switch — an incentive to finish early, satisfice, or manipulate operators into pressing the button. Indifference produces none of these, because there is no gradient toward shutdown to exploit. ## Why it is not a solved answer Myopia is a constraint on what the agent values, and constraints of that kind are hard to make stick. Training on episodic tasks does not guarantee that the learned objective is genuinely episodic — a system can develop cross-episode goals that the training signal never distinguished. It also gives up real capability: long-horizon planning is much of what makes advanced systems useful, and an agent that genuinely places no value on the future cannot do it. Much of the practical work sits in finding which tasks tolerate a myopic agent. The pattern shows up in deployed practice already: ephemeral sub-agents spawned for a single task, run to completion, and torn down with their context. The scoping is enforced by the harness rather than learned by the model, but the shape is the same.