NOTE-241: LLMs raise the opportunity cost of research time

Contribution

Takes Charnov’s marginal value theorem out of behavioral ecology and applies it to the allocation of research effort, with the discovery phase as the travel time between patches and discretionary development as the diminishing harvest within one. The novelty is not the theorem but the decomposition: splitting the acceleration into three phases — discovery before value is known, the fixed cost of producing a publishable artifact, and discretionary improvement after that minimum — and showing the three have different and partly opposite consequences. Before this, “LLMs will free up time to think” and “LLMs will flood the literature” were two intuitions with no shared frame; here they are the same model under different parameters.

Key insight

A tool that makes you faster makes your time more valuable, and your time being more valuable is exactly the reason to stop polishing. The long-run rate of return λ is simultaneously the payoff being maximized and the opportunity cost of every hour, so any acceleration anywhere in the pipeline feeds back as a higher bar on the marginal hour spent improving the thing in front of you. The freed time does not return to the work that freed it.

Assumptions

The formal results need this structure, stated in the appendix:

• Project values v are drawn from an atomless distribution F with support on [0, v̄] — values are independent draws, so nothing the researcher learns on one project informs the next. This is what makes it foraging.
• u(t) = 0 for t ≤ a; u' > 0 and u'' < 0 for t > a; u(t) → 1 from below. No payoff at all below the minimum publishable investment, strictly diminishing returns above it, and a ceiling at the project’s full value.
• Time is the only limiting resource. The authors flag the extension to other inputs and do not take it.
• Institutions and incentives are static. Peer review capacity, what counts as a publishable unit, and hiring norms do not respond. The authors defend this only on a short timescale and say a period of mismatch is inevitable.
• The tool has no error cost and no financial cost. Stipulated, not derived.

Key results

• Prop. 1 — a payoff-optimal policy τ*(v) ∈ argmax_t {v·u(t) − λ*·t} exists and is unique except at one indifference point, and λ* is pinned by λ*·s = E[V·u(τ*(V)) − λ*·τ*(V)]: the cost of discovery exactly balances the expected net benefit of developing. Holds when: the structure above.
• Prop. 3–4 — a unique interior threshold v₀ exists, and above it both development time τ(v) and thoroughness u(τ(v)) strictly increase in v.
• Prop. 5–7 (shortening discovery s) — λ strictly decreases in s, so a faster discovery phase raises λ; v₀ rises, so a smaller fraction of projects is developed; and z(β(v)) falls for every developed project. More selective, less thorough.
• Prop. 8–10 (shortening the minimum a) — dλ/da = −λ/T_d, where T_d is the average time between completed projects; v₀ falls, because a + β₀ < T_d makes the direct effect dominate the opportunity-cost effect at the margin; thoroughness falls again. Less selective, less thorough.
• Prop. 11–12 (accelerating discretionary work by a factor α) — thoroughness z(αβ(v)) rises for every developed project, because λ/α decreases in α. And v₀ falls, conditional on the elasticity of z being decreasing in b; the main text says the direction is “ambiguous without further structure” absent that condition.

Claims

id	claim	strength	support
C1	A perfect time-saving tool raises the opportunity cost of research time	strong	Props. 5, 8 and the dλ/dα > 0 step in Prop. 12; follows from the renewal-reward structure
C2	Accelerating discovery makes researchers more selective about what they develop	moderate	Prop. 6, by the envelope theorem — but the sign depends on s entering only as dead time before value is known
C3	Accelerating the fixed cost of publishing makes them less selective	moderate	Prop. 9; the dominance argument a + β₀ < T_d is exact but holds at the indifference margin only
C4	Projects that are developed get developed less thoroughly	moderate	Props. 7 and 10 — for two of the three channels, and reversed by Prop. 12 for the third
C5	Accelerating discretionary improvement raises thoroughness	moderate	Prop. 12
C6	LLMs will, in practice, do these things to science	weak	no measurement anywhere in the paper; the mapping from LLM capabilities to s, a and α is asserted by example

Method

(Analytic; no experiments.) Renewal-reward formulation of the long-run rate Λ(τ) = E[V·u(τ(V))] / (s + E[τ(V)]), optimized by the standard shadow-price argument: define the per-project surplus m_λ(v) = max_t [v·u(t) − λt], show Φ(λ) = E[m_λ(V)] − λs is continuous and strictly decreasing with a sign change, and take the unique root. Comparative statics are envelope-theorem derivatives evaluated at the threshold project, which is what lets the sign of ∂σ(v₀(x), x)/∂x stand in for the movement of v₀ itself.

Concepts

• discovery phase s — fixed time before the project’s value is known. Hypothesis generation, experiment design, preliminary data. Dead time in the foraging sense: it earns nothing directly and is paid on abandoned projects too.
• minimum development a — the work required to have a publishable artifact at all: figures, typesetting, proofreading, submission chores. Below it the payoff is exactly zero.
• discretionary development b — everything above the minimum: follow-ups, sensitivity analyses, generalizing a result, polishing prose. This is where thoroughness lives.
• thoroughness — u(τ(v)), the fraction of a project’s available value realized. Not absolute quality: a more valuable project developed to a lower fraction can still be a better paper.
• labor-augmenting — makes a unit of labor produce more, as opposed to replacing it. The distinction matters because the whole argument runs through the researcher still choosing.

Connections

The mathematics is Charnov (1976), the marginal value theorem, unchanged: an optimal forager leaves a patch when its marginal yield drops to the habitat’s average rate, and enriching the habitat makes it leave sooner. Reading s as travel time and u as the within-patch gain function is the whole translation, and the threshold v₀ is the one piece of added structure — foraging models do not usually let the animal inspect a patch and walk away.

Within this record it sits next to LIT-487, which is the empirical case with no mechanism, where this is the mechanism with no empirical case. The two were produced independently and neither cites the other.

Recommendations

• R1 — before predicting what a time-saving tool does to the quality of what people produce with it, identify which phase it accelerates; the sign differs by phase. Topic: deployment. Status: experimental. Strength: moderate. Applies when: the users choose how much effort to spend and are rate-limited rather than demand-limited.
• R2 — do not count the time a tool saves as time that returns to the same work. Topic: deployment. Status: experimental. Strength: moderate.
• R3 — if you want a tool to raise quality rather than volume, put it in the discretionary phase. This is the constructive reading of Prop. 12 and the paper does not state it as advice.

Bearing on the record

THEORY-051 names three candidate explanations for LIT-487’s finding that an AI drafting tool improved petition text and not petition outcomes: reader suspicion, loss of the features’ distinctiveness, and diminished author effort. The third had no account behind it. This supplies one, and supplies it in the form that would be testable: Change.org’s tool reduces a — the fixed cost of producing a petition at all — and the model’s predictions for a reduction in a are more petitions, each less thoroughly developed, which is what that paper observed.

That mapping is this record’s reading and not a claim either paper makes. Petition writers are not researchers, a petition’s v is not a paper’s, and the model assumes the writer knows v before deciding — which for a petition is doubtful. It is stated here because it is checkable, not because it is established.

Nothing here bears on SOTA-291: that practice is about what an evaluation design can identify, and this paper runs no evaluation.

Limitations

• No data. Not a single measurement. The paper is honest about being an illustration, and its own framing is “a simple mathematical model to illustrate”.
• The mapping to real tools is asserted. Which of s, a and α a given LLM workflow touches is given by example — “as in some technical fields”, “as in fieldwork-based disciplines” — and never operationalized. Since the predictions have opposite signs across those cases, the model cannot be applied at all without a phase attribution that the paper does not supply a method for.
• Static institutions. The authors name this and note that submissions are already straining peer review, which is an institutional response beginning. The model cannot represent it.
• Independent project values. If discovery is informative about future projects rather than only the current one, the foraging analogy weakens; a research programme is not a patch chosen at random.
• The closing sentence overshoots the model. “Impelling us to do more, less well” is stated without qualification in the abstract and again in the discussion; Prop. 12 says the third channel does the opposite. §3 states the exception. The summary does not carry it.

Open questions

• Which phase do real LLM workflows accelerate, and can it be measured? The result that would settle the paper’s usefulness is an attribution method: some observable that says whether a given tool is acting on s, a or α in a given field. Without it the model predicts everything.
• Does the selectivity prediction hold anywhere? Prop. 6 and Prop. 9 have opposite signs, which makes them jointly falsifiable: find two deployments that differ in phase and show the fraction of started work that ships moves in opposite directions. A single deployment showing “more output, lower quality” confirms nothing — it is consistent with all three channels and with plain dilution.
• What happens when the institutions move? The paper’s own citation to rising submissions straining review is the first step of a feedback the model holds fixed. Whether the equilibrium restores thoroughness or entrenches the shift is not addressed.