I examine what it means for AI assistance to save time and contrast two perspectives: the stopwatch view, which equates saved time with the difference shown on a clock, and the ledger view, which treats saved time as a counterfactual that must be read alongside the defects that slip through the work. I argue for the latter.
First, I define, without assuming any model, two quantities—time saved and excess escaped defects—and a decision metric built from them: the break‑even defect cost, denoted $C_{be}$.
Then I present a baseline checking model: as long as the model is not good enough to forgo checking, a calibrated worker’s escaped‑defect rate is independent of model quality; checking time grows with the logarithm of the model’s error rate; and over‑trust buys minutes on the clock at a larger total minute cost.
I discuss what randomized studies can identify and replicate METR’s developer study from its public data. The 246 issues support a slowdown of 19% or 24% depending on the estimand (34% without adjusting for a chance imbalance), and AI helped on somewhere between 3% and 82% of them.
Finally, I restate the published record, replacing point assumptions with explicit bounds. In GDPval’s review‑every‑output workflow, review time alone caps GPT‑5’s speedup at 3.6×; consultants outside the frontier were faster but more often wrong.
I also test two model predictions in a small pre‑specified experiment on AI reviewers checking AI‑drafted briefs built from SEC filings collected after the models’ training data. When told the drafter’s true error rate, the reviewer checked more as it rose, but in an all‑or‑nothing fashion, switching near the prompt’s break‑even cost. When told the drafter erred ten times less often than it actually did, the reviewer checked 7 fewer of 9 figures and let a wrong figure through in 22.5% of briefs instead of none—faster and worse, entirely due to reduced effort.
Review