Four Questions to Ask About Any AI News

In September 2026, Dario Amodei’s essay proposed a three-step plan to pace frontier AI. Several commentaries warned that safety compliance could harden into an incumbent moat, while a counter-analysis argued that regulatory capture alone was an incomplete diagnosis.¹²³⁴ The clash is a useful doorway because it tempts readers to choose a camp before separating the claim, the intervention, and the institution that would enforce it.

A reusable test asks four questions. The table applies each question twice, first to the technology or danger being described and then to the proposed response.

Rate of change is about mismatch, not speed in isolation. For the claim, identify exactly what is said to be accelerating and what evidence could keep pace with it. For the remedy, identify how rapidly it would alter institutions, obligations, or control. This prevents urgency from becoming a substitute for specifying what has changed.

Reversibility deserves its own question rather than being folded into likelihood. Toby Ord argues that extinction-level or civilization-ending outcomes deserve additional weight partly because they cannot be reversed, separately from any estimate of how likely they are.⁵ The economics of irreversible investment similarly identifies value in preserving the option to wait when uncertainty and irreversibility occur together.⁶ Neither argument establishes that waiting is always correct in an AI case. They establish reversibility as a distinct axis of judgment.

Verification asks what observation could distinguish observance from assertion. A promise about intent is not the same kind of evidence as inspectable conduct. Concentration asks who receives authority and who can absorb the compliance burden. One moat analysis, for example, argues that requiring a company’s existing safety apparatus can burden lower-cost competitors disproportionately.² Asking about concentration tests institutional effects without assuming that every concentration of power is illegitimate.

This framework cannot tell the reader whether a forecast is true, whether Amodei is sincere, or which intervention is best. It is not a scorecard where favorable answers prove a case. Apply it symmetrically to the warning and the remedy. How fast is the relevant change? What is reversible? What can be verified? Where does power concentrate? Those questions are the conclusion, not a verdict.