The saved-hour fallacy: a model in which better AI makes each paper worse

The load-bearing part of the model is which stage of a project has someone enforcing it. Pushing a project forward splits into mandatory work, such as making figures, formatting and submitting, and voluntary work, such as extra experiments, deeper analysis or polishing the prose ^[5]. A venue enforces the first. Nothing enforces the second except the researcher's own judgement, and that is precisely what gives way once the next project starts to look cheap ^[6].

That asymmetry is why the direction the tooling points matters more than how much time it saves. Accelerate early triage and researchers get choosier, because abandoning an idea costs less, but the survivors still get thinner treatment, since the freed hours pay better in a new project ^[7]. Accelerate the write-up and the bar for "worth submitting" drops, so marginal projects go out and the average paper is shallower ^[8]. Only when the tool speeds the voluntary stage does the saving land where the corner-cutting already happens ^[9].

Hold that taxonomy against the field evidence in the same write-up. OpenAI's report on eight case studies claims up to 60x speedups when rewriting research software, with the bottleneck moving to validation and long-term maintenance ^[11]. Validation is voluntary-side work in the paper's sense: no reviewer form demands it, and no deadline is missed by skipping it. So the measured win sits in the mandatory column and pushes load into the column the model says gets sacrificed first. That is scenario two with a compiler attached.

The METR result is the sharper test, because it removes the need for the time saving to exist. Experienced open-source developers using AI tools took 19 percent longer while reporting they felt 24 percent faster ^[12], a 43 point spread between measured and felt speed ^[13]. The write-up notes the behavioural point directly: perceived savings do not have to be real to change how people allocate effort ^[14]. An incentive that keys off belief is not fixable by making the model better, and it is not detectable by asking users whether the tool helped.

None of this depends on hallucination, cost or model error. The authors deliberately treat language models as error-free and financially negligible in order to isolate the time effect ^[3], and the pessimistic result survives that generosity ^[4]. Their own summary is that as a labor-augmenting technology, LLMs "increase the opportunity cost of our time, impelling us to do more, less well" rather than "the same amount, better" ^[10]. Read as a procurement note rather than a lament, it says something narrower and more usable: a tool that shortens the parts of the job someone else checks buys throughput, and only a tool that shortens the parts nobody checks can buy care.