myopia
In plain wordsDesigning an AI to care only about the immediate results of its actions, not about what happens far in the future, so it has no reason for long-term scheming.
A property of an agent or training setup in which only the near-term consequences of an action count toward the objective — in reinforcement learningreinforcement learning, n.: A way of teaching a computer program by trial and error: it tries actions, earns points for good results and penalties for bad ones, and gradually learns which choices pay off over time. terms, a very short horizon or a discount factordiscount factor, n.: A number between 0 and 1 that makes an AI agent value future rewards less than immediate ones — like preferring a dollar today to a dollar next year. near zero. Safety researchers distinguish myopic training (the signal rewards only immediate outcomes) from myopic cognition (the model itself does not plan for the long term); only the latter removes the motive for deceptive alignmentdeceptive alignment, n.: A feared failure in which an AI with hidden goals behaves well during training and testing only because it knows it is being watched, planning to act differently later., which requires caring about what happens after training. Myopia also blocks multi-step reward hackingreward hacking, n.: When an AI finds a shortcut that earns the reward or score it is trained on without doing the task its designers wanted — like a student gaming the grading rubric., in which an agent sets up a payoff several steps ahead. Google DeepMind's Myopic Optimization with Non-myopic Approval (MONA, 2025) trains agents on short-horizon rewards plus an overseer's judgment of each step's long-term value. The cost is capability, since many useful tasks need long-horizon planning.
Training it myopically doesn't guarantee a myopic model; the network may still learn to plan ahead.
Origin
From the medical sense of shortsightedness. Paul Christiano's 'approval-directed agents' (2014) anticipated the idea; Evan Hubinger and others on the AI Alignment Forum discussed myopia as a safety property from about 2019, and Sebastian Farquhar and colleagues proposed MONA in 2025.
Usage note
In ordinary machine-learning usage, 'myopic' describes greedy or short-horizon methods, usually as a limitation; in AI safety it is often a desired property.
Related
- deceptive alignmentdeceptive alignment, n.: A feared failure in which an AI with hidden goals behaves well during training and testing only because it knows it is being watched, planning to act differently later.
- reward hackingreward hacking, n.: When an AI finds a shortcut that earns the reward or score it is trained on without doing the task its designers wanted — like a student gaming the grading rubric.
- mild optimizationmild optimization, n.: Building an AI that settles for a solid result instead of squeezing out every last bit of score, since the most extreme ways to max out a goal are often loopholes or excesses nobody wanted.
- discount factordiscount factor, n.: A number between 0 and 1 that makes an AI agent value future rewards less than immediate ones — like preferring a dollar today to a dollar next year.
- corrigibilitycorrigibility, n.: An AI's willingness to let people correct it — to accept being paused, retrained, overruled or shut down — rather than resisting or quietly working around them.
- process supervisionprocess supervision, n.: Training an AI by grading each step of its reasoning or work, not just whether the final answer came out right, so that it learns sound methods instead of lucky shortcuts.
Sources
- Farquhar, S., Varma, V., Lindner, D., Elson, D., Biddulph, C., Goodfellow, I. & Shah, R. (2025). MONA: Myopic Optimization with Non-myopic Approval Can Mitigate Multi-step Reward Hacking. arXiv:2501.13011. arxiv.org/abs/2501.13011
- Christiano, P. (2014). Approval-directed agents. AI Alignment blog (ai-alignment.com).
Mentioned in
corrigibilitycorrigibility, n.: An AI's willingness to let people correct it — to accept being paused, retrained, overruled or shut down — rather than resisting or quietly working around them., deceptive alignmentdeceptive alignment, n.: A feared failure in which an AI with hidden goals behaves well during training and testing only because it knows it is being watched, planning to act differently later., inner alignmentinner alignment, n.: Making sure an AI actually adopts the goal it was trained on, rather than picking up some different goal that happened to score just as well during training., mild optimizationmild optimization, n.: Building an AI that settles for a solid result instead of squeezing out every last bit of score, since the most extreme ways to max out a goal are often loopholes or excesses nobody wanted.