LIT-492: The unintended consequences of large language models as a labor-augmenting technology in science

Duede, Gross, Crockett and Bergstrom (2026) — ARXIV-2607.17397, read as NOTE-241.

Key takeaways

• The tool is stipulated to be perfect. Hallucination, fabricated citations, sycophancy and deskilling are listed in the introduction and then set aside: the model gives the LLM “as imagined by their most enthusiastic proponents”, reducing time cost without adding error, at negligible price. Every consequence below is therefore a consequence of speed, not of defect — which is what makes the paper hard to answer by improving the tool.
• Charnov’s marginal value theorem, with a threshold. A researcher spends a fixed discovery phase s learning a project’s value v, then either abandons it or develops it for t = a + b, where a is the minimum needed to publish at all and b is discretionary improvement. Payoff is v·u(t) with u zero below a and concave above it. The optimal policy maximizes the long-run rate λ, and is characterized by two things: a threshold v₀ below which projects are abandoned, and a first-order condition v·u'(τ(v)) = λ for the rest.
• λ is the opportunity cost of time, and every acceleration raises it. That single quantity drives the results. Raise it and the first-order condition binds sooner, so each developed project gets less discretionary work.
• Selectivity splits by phase, and the split is the paper’s real content. Shortening discovery s raises v₀ (Prop. 6): returning to the drawing board got cheaper, so fewer projects clear the bar. Shortening the fixed publication cost a lowers v₀ (Prop. 9): the opposing effect on the opportunity cost is always dominated at the margin of indifference. More published, or less, depending on which phase the tool touches.
• Thoroughness falls in two channels and rises in the third. Props. 7 and 10 give the decline for s and a. Prop. 12 gives the increase when the tool accelerates discretionary development itself, because that raises the real value of b faster than it raises λ. Prop. 11 adds that this third channel also lowers v₀, but only under a condition — that the elasticity of the benefit curve is decreasing — which the main text acknowledges as “ambiguous without further structure”.

Standing in the anthology

Kept as a mechanism the record was carrying an empirical shadow of and no account for. LIT-487 measured a writing tool that made its users’ text better on every feature and their outcomes no better, and THEORY-051 named three candidate explanations for that; one of them was diminished author effort, and nothing in the record said why effort would diminish. This supplies a reason that does not require the tool to be bad.

It is a model with no data. Nothing here is measured, the institutions are held static by assumption, and the authors say so. What it buys is a sign prediction that depends on an identifiable property of the tool, which is more than “more output, worse quality” — that conclusion is compatible with almost anything, and this says when to expect the opposite.

One thing to carry when citing it. The closing line — “impelling us to do more, less well” — is the most quotable sentence in the paper and is the one its own Proposition 12 contradicts for one of the three channels. The analysis section states the exception plainly; the abstract and the discussion generalize past it. DP-010 in both directions, in a single paper.