Low-impact agents
Concrete Problems in AI Safety — inherited
An informally-discussed prior concept (Armstrong 2012) naming agents designed to minimize their impact on the environment beyond their intended task.
In 2012, four years before Concrete Problems, Stuart Armstrong sketched the idea of AI systems built to disturb the world as little as possible beyond their assigned task. Low-impact agents is the name that sketch gave the field.
That early sketch named the side-effect problem before the paper existed, and thinking about it shifted over time, from measuring the change an agent actually causes to limiting the power it merely holds. Penalize influence, the proposal that carried the idea forward, is the natural next stop.
Naming the problem before the paper existed
Low-impact agents is Stuart Armstrong's 2012 informal proposal for AI systems deliberately designed to minimize their effect on the environment beyond whatever their intended task strictly requires. It predates Concrete Problems by four years and belongs to an earlier, more philosophically oriented wave of AI-safety writing that discussed such constraints discursively rather than as a formal optimization problem; Concrete Problems cites it as the closest prior name for what it calls Negative side effects (avoiding) (concrete-problems, §"3 Avoiding Negative Side Effects", p. 4).
From realized change to potential power
Most readings of "low impact" are naturally about realized change: an agent should not alter the world more than its task demands. The edge Low-impact agents — gets reinterpreted as potential power by → Penalize influence shows how the paper's own Penalize influence proposal quietly relocates that target. Its motivating example is a cleaning robot that should avoid bringing "a bucket of water into a room full of sensitive electronics, even if it never intends to use the water in that room" (concrete-problems, §"Penalize Influence:", pp. 5-6): nothing has changed yet, no low-impact violation has occurred under the ordinary reading, and yet the agent has acquired latent power to cause harm. This matters because an agent could satisfy every conventional test of "low impact" while still accumulating exactly this kind of dormant capability, which is why the corpus treats penalize-influence as a genuine extension of the low-impact target rather than merely one more implementation of it. The concept sits in the Limiting potential influence instead of measuring realized change theme, the A safety mechanism built by patching together other people's guarantees connective theme, and the Bounding what an agent can change or explore, not just what it's told to want supertheme, functioning as the conceptual seed from which the paper's more precise, information-theoretic treatment of influence grows.