The paper draws an explicit line around its own territory, and names what falls outside
Some papers only imply their own boundaries; this one draws them twice and names exactly what falls outside. Flagging a failure as exceeding your own frame, in the very paragraph that introduces it, is an odd move, but that's how the shutdown problem gets treated here.
A narrowing away from long-term AI-risk framing opens the page, followed by that shutdown-problem hedge and wireheading marked as narrower than reward hacking in general. Privacy's quiet contrast with security research, a one-line definitional gap, closes it.
Concrete Problems repeatedly draws a boundary around its own scope and is explicit about what sits on the other side of it. Accidents in ML reframes away from long-term AI risk framing, a deliberate methodological narrowing toward precision over speculation rather than a rejection of the broader concern. Multi-agent approaches flags the shutdown problem as exceeding its own frame, hedging that resisting correction is a structurally different failure than an ordinary externality even though the same section introduces it. Wireheading is narrower than reward hacking, a taxonomy-scoping move the paper needed once it found failures, like Goodhart's-law-style proxy gaming, that produce reward-hacking behavior without any tampering step. And privacy needs no adversary unlike security ML, a one-line definitional gap that tracks a real methodological difference between the two adjacent research areas the paper places itself next to.