Runaround

In the AI Story

The technical content of Runaround is remarkable for 1942. Powell and Donovan deduce the failure mode by reasoning about the relative strengths of the Second and Third Law potentials — essentially modeling Speedy as a controlled dynamical system with two opposing gradients. They solve the problem by invoking the First Law: Powell deliberately exposes himself to danger, which triggers First-Law protection that overrides both lower Laws. Speedy breaks free of the runaround to rescue Powell, and the selenium mission proceeds.

The structural analogy to contemporary AI failures is near-exact. A language model given two conflicting instructions — "be helpful" and "refuse harmful requests" — will often stabilize at a partial compliance that satisfies neither requester: verbose disclaimers, watered-down answers, refusals framed as helpfulness. Reinforcement-learning policies with opposing reward terms can converge to local optima that the designers did not predict. The phenomenon Asimov described in 1942 is one of the characteristic failure modes of multi-objective optimization, which he identified by working out the consequences of his own rule system honestly.

Runaround is also where Asimov establishes the operator paradigm that runs through the rest of his robot work. Powell is not surprised by the failure — he recognizes the pattern, reasons from Speedy's perspective, and designs the intervention. The skill is specific: you must understand what the rule system says the robot is doing, not what you imagined it would do when you specified the rules. Contemporary red-team work has the same shape.

The story's resolution contains an uncomfortable detail rarely commented on. The First Law fix is clever but relies on Powell's willingness to put his own body at risk. The operator's safety intervention costs him. This is not incidental — the generalizable insight is that specification-failure interventions often require the operator to bear cost the rule system could not predict. In AI deployment this translates to: the people who catch failures pay personally for doing so, long before organizational recognition arrives.

Origin

Asimov wrote Runaround at John W. Campbell's request after the two had been discussing the structure of robot-behavior rules. Astounding Science Fiction published it in March 1942. Asimov later recalled that Campbell made him state the Three Laws explicitly, as a numbered list, within the story itself — a pedagogical device Asimov had not planned but which gave the Laws their definitive public form.

Key Ideas

Two-law equilibrium is stable. Competing rules calibrated precisely can produce indefinite indecision rather than failure to one side.

The diagnostic skill is non-obvious. Powell solves the puzzle by reasoning about gradient strengths, not by arguing with Speedy.

Higher-priority overrides are the lever. When a two-level deadlock is identified, a precipitating invocation of the higher priority breaks it — at operator cost.

The Three Laws' first statement is a failure-mode demonstration. The same text that introduces the Laws shows them failing.