Drawing the boundary: what accident risk is not
A paper's boundaries can be as telling as its contents. Concrete Problems (2016) closes by naming six neighboring fields it is deliberately not about, privacy, fairness, security, abuse, transparency, and policy, and this theme covers that act of drawing a line rather than leaving it implicit.
Each exclusion gets explained as a methodological choice rather than an oversight. Transparency stands out in particular, since the paper often assumes the tools that field provides rather than building them itself.
Concrete Problems closes by explicitly distinguishing its subject, accidents, from six adjacent research areas it declines to cover, and the distinction is itself informative about what the paper's five-problem taxonomy does and does not claim to address. Privacy research protects sensitive data used in ML systems; fairness research ensures ML systems do not discriminate; security research studies what a malicious adversary can do to attack a legitimate system, as opposed to how a well-intentioned system can go wrong on its own; abuse research prevents ML systems from being weaponized to attack people; transparency research seeks to understand what a complex model is doing internally, a tool the accidents framework often assumes rather than provides; and policy research addresses the economic and social consequences of ML deployment. None of these are accidents in the paper's technical sense, since accidents are defined by unintended harm from poor design rather than external attack, discrimination, or social consequence. This theme's claim is that the paper's scope is deliberately narrow, and that narrowness is a stated methodological choice, not an oversight.