Where Concrete Problems draws its own lines
A paper this deliberate about its remedies turns out to be just as deliberate about its limits. Concrete Problems draws its own boundary twice, once at the start against superintelligence framing, once at the end against six neighboring fields it declines to touch.
Between those two boundary-drawing moments sits a smaller point worth noticing: the same toolkit built for formal verification also ends up rehabilitating Asimov's first law, turning a seventy-year-old slogan into something closer to a checkable property.
Before and after its five-problem taxonomy, Concrete Problems is deliberate about what it is and is not claiming. Framing safety as accidents, not superintelligence stakes out the opening move, defining its subject as accidents in machine learning systems rather than the pre-existing long-term or superintelligence framing, and rather than fictional shorthand like Asimov's first law, using the cleaning-robot example to keep discussion mundane and tractable. Drawing the boundary: what accident risk is not performs the same move at the paper's close, distinguishing accidents from six adjacent fields, privacy, fairness, security, abuse, transparency, policy, that the paper declines to cover. Formal verification and control theory recur as safety building blocks connects to both: Asimov's first law reappears here as an early, informal attempt at exactly the kind of machine-checkable safety property the paper's formal-methods toolkit later makes rigorous. Together these three themes show a paper being explicit about its own boundaries, twice, and about the maturity of the tools available to it.