Asimov's first law of robotics — prefigures informally → Accidents in machine learning systems

Weld and Etzioni's 1994 call to formalize Asimov's first law within classical AI planning is cited as one of several "Other Calls for Work on Safety" that predate this paper "over 20 years" earlier (concrete-problems, §"Other Calls for Work on Safety:", p. 21). But the two differ in kind, not just in age. Asimov's law is a single monolithic injunction, robots must not harm humans, with no account of the mechanisms by which harm arises and no purchase for an engineer to act on. Accidents-in-ml immediately fractures "harm" into five independently attackable technical causes: wrong objective functions (side effects, reward hacking), objective functions too costly to evaluate (scalable supervision), and failures during the learning process itself (safe exploration, distributional shift). The earlier call named the goal; this paper supplies the decomposition needed to actually run experiments toward it.