BuildNot yet confirmed elsewhere1 publisher3 min readPublished
The saved-hour fallacy: a model in which better AI makes each paper worse
A theoretical paper assumes error-free, near-free language models and still finds that in two of three deployment scenarios, researchers respond by doing more work with less care.
The Engineer · Build desk
What happened
- Researchers from Princeton, the University of Washington and other institutions published a theoretical economics paper on how AI time savings change researcher behaviour.
- The model borrows optimal foraging theory from behavioral ecology to simulate how scientists spread effort across competing projects.
- The authors assume language models make no errors and cost almost nothing, so the only variable left is time.
- Under those assumptions, two of three deployment scenarios reduce per-project thoroughness, and only one raises it.
Why it matters
- constraint Quality gains are available only from tools aimed at the stage nobody audits, which means most drafting, formatting and triage assistance can buy volume but cannot buy care.
- decision Where a lab points its tooling becomes a quality decision rather than a speed decision, and the model says technical and fieldwork groups face opposite failure modes.
- exposure The bill for the publishing scenario lands on unpaid reviewers, who absorb more submissions at lower average depth; screening is the one layer the tools do not accelerate.
- contradiction The incentive runs on believed time savings, so a tool that measurably slows its users can still cut the effort they put into each project, and no amount of model improvement addresses that.
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 [15]. The write-up notes the behavioural point directly: perceived savings do not have to be real to change how people allocate effort [13]. 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.
What to watch
- Whether any vendor ships tooling aimed at the voluntary deep-dive stage, the only place the model says quality rises, rather than at drafting and formatting.
- Whether submission volumes diverge between fieldwork-heavy and technical fields, which is the split the model predicts.
- Whether further measurement replicates the gap between felt and actual speed as the tools change.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence42
- Adoption28
- Hype gap+18
- Incentives46
- Confidence38
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
A theoretical economics paper by researchers from Princeton, the University of Washington and other institutions argues that because AI saves time, researchers will spend less effort on each project rather than more.
- [2]
The model is built on optimal foraging theory from behavioral ecology, a framework describing how organisms allocate effort across competing opportunities, adapted to simulate how researchers distribute labor across projects when LLMs shorten different phases.
- [3]
The authors deliberately idealize LLMs as tools that cut time costs without introducing errors and at negligible financial cost, to isolate the pure effect of time savings.
- [4]
In two of the paper's three scenarios thoroughness drops; only one scenario improves quality.
- [5]
In the model a project has two phases: the researcher checks whether an idea is viable, then either abandons it or pushes forward; pushing forward has a mandatory part (creating figures, formatting text, submitting) and a voluntary part (running extra experiments, deeper analysis, polishing prose).
- [6]
The authors argue the voluntary part is what gets sacrificed when time becomes scarce, because polishing an already publishable paper stops making sense once time is more valuable.
- [7]
Scenario one: AI helps evaluate early ideas, so researchers become pickier because starting over is cheaper, but even the surviving projects get less thorough treatment because the saved time is better spent launching something new; the authors say this pattern is typical of technical fields.
- [8]
Scenario two: AI speeds writing, formatting and analysis, so weaker projects become worth pursuing, more papers enter circulation and each is shallower; the authors say this pattern is typical of fieldwork-based disciplines.
- [9]
Scenario three is the only one where quality improves: AI speeds the voluntary deep-dive phase, the exact stage where researchers have always cut corners under time pressure, so the savings translate into more thorough work.
- [10]
The authors write 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".
- [11]
A field report from OpenAI covering eight scientific case studies found up to 60x speedups when rewriting research software, but the bottleneck shifted from coding to validation and long-term maintenance.
- [12]
A METR study found experienced open-source developers using AI tools took 19 percent longer to finish tasks even though they felt 24 percent faster.
- [13]
The write-up notes that perceived time savings do not have to be real to change behavior.
- [14]
Sakana AI's "AI Scientist-v2" pushed a fully AI-generated paper through an ICLR workshop, citation errors included.
- [15]
The gap between measured and perceived speed in the METR result is 43 percentage points.
- [16]
In fields where LLMs speed up writing, submissions are already climbing fast and straining an already overloaded peer review system.
Sources
1 independent publisher whose own reporting we read for this story.
- the-decoder.comAI could make scientists do more work less well, not less work better, study argues
1 article · August 23, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- AI for ScienceFollow
- Economics of labor-augmenting AIFollow
- Scientific publishing integrityFollow
- AI and research productivityFollow
- Developer Productivity MeasurementFollow