Game theory of academic effort and hiring in the age of AI

A research paper used to be decent evidence that its authors understood their field. When AI can produce plausible theoretical work cheaply, that signal weakens, and the consequence runs through hiring. Pal and Park model the academic job market as a two-sided game between early-career applicants choosing how deeply to study and recruiters choosing what evidence to require, and classify the five pure-strategy equilibria that can arise. Their central result is a coordination trap: valuing understanding does not by itself create an incentive to acquire it, because when no applicant studies deeply a passed interview carries no information and is correctly ignored. They also identify a reverse-sorting equilibrium in which medium-potential applicants study deeply while high-potential ones rely on AI, and learning simulations suggest populations often settle where nobody studies deeply even when deep study could be sustained.

Key Findings

• The signal has decoupled from the skill it proxied. Once AI makes strong-looking output cheap, a well-written paper no longer reliably sorts candidates by depth of understanding, and recruiters who correctly interpret the signal can still end up rejecting everyone.
• Five equilibria describe the baseline market. E1 has all strategic applicants shallow and nobody hired; E2 has only high-potential applicants deep; E3 has all strategic applicants deep; E4 is reverse sorting; E5 again has all shallow with a different hiring rule.
• Valuing understanding is not sufficient. In a market where everyone studies shallowly, a pass looks identical to a lucky shallow pass, so deep study earns no reward even when recruiters put a high value on understanding.
• Institutional expectations select among equilibria. The same valuations and test accuracy can support either a deep-study market or a shallow one; what differs is each side's belief about the other, and each belief is confirmed by the behavior it produces.
• Reverse sorting is possible. With AI amplifying output for high-potential applicants, high-potential candidates can rationally go shallow while medium-potential candidates go deep, because the output signal is cheaper for the former to fake.
• Coordination is hard to reach in simulation. In the rare-exploration limit, where only one population changes at a time, populations often settle in the no-deep-study equilibrium, and faster exploration that lets several departments move at once may help.

Modeling a two-sided market

The model gives applicants two strategies, deep study or shallow work with AI, and recruiters a hiring rule conditioned on observed output and on an interview that probes understanding with accuracy τ. The interview is the only noisy window onto depth, so the analysis asks what τ, the value recruiters place on output (A) and on understanding (B), and the cost of study jointly imply for equilibrium behavior. Equilibria are characterized as pure-strategy perfect Bayesian equilibria, and the paper reports the recruiter's beliefs at each, which matters because the beliefs are what make a hiring rule self-fulfilling.

An extended "super-output" version adds a third output level produced by high-potential applicants using AI, because the practical worry is not that AI produces an average paper but that it lets the strongest candidates produce more of them. That extension is where reverse sorting becomes available, and where the model speaks most directly to departments worried about losing their ability to identify depth.

The coordination trap, in plain terms

The trap is stated as a two-step argument. Recruiters know that if everyone studies shallowly, then everyone who passes the test got through by luck, so a pass carries no information and, if output alone does not justify hiring, recruiters reject all applicants. Now imagine one applicant studies deeply: they pass, but a pass looks exactly like the lucky shallow passes, so the recruiter still cannot distinguish them and rejects them too. Deep study brings no reward, no applicant changes course, and the market sits where it began.

This is why the paper argues that the levers are how understanding is rewarded and how reliably it is assessed, rather than exhortation. It also explains why the same fundamentals support two outcomes: expectations about what a pass means determine whether a pass is read as evidence, and those expectations are sustained by behavior rather than by evidence.

What departments can actually change

The practical section is modest and worth taking literally: two priorities, rewarding understanding and assessing it accurately. Improving τ has an unambiguous role in the model, since a more accurate interview makes a pass informative; but the equilibrium analysis warns that an accuracy improvement alone may not move a market already settled in the shallow basin. The learning simulations add a second lever, which is the pace of institutional change: when single departments can only shift one at a time and are penalized for moving alone, populations stay stuck, whereas coordinated moves by several departments relax that constraint. The paper's own examples of real responses, less weight on dissertation text and more on in-person defense, at most three works reviewed per applicant with chalk talks, and capped preprint submissions, are all moves in this direction, and the model says why they need to be made together rather than one candidate at a time.

What this means for practice

• Assess depth in person, and say so in advance. An interview that probes understanding is the model's only informative channel, so publish that it is decisive rather than an optional extra.
• Move as a group of departments. Because coordination traps survive unilateral reform, hiring consortia and shared interview protocols change behavior in a way that a single department's stricter rubric may not.
• Treat output volume as evidence about AI access, not only about ability. Where AI amplifies output unevenly, a longer publication list can mark the candidates who could afford the tool.
• For administrators, separate the reward for understanding from the reward for output. Valuing understanding while paying only for volume reproduces the equilibrium the paper describes.

Limitations

• The model is a stylized two-type applicant market with one interviewer signal, so real hiring committees, subfield variation and multi-stage searches are outside it.
• Learning simulations use a specific parameterization (N = 30; interview accuracy τ ∈ {0.5, 0.9, 0.99}; a 101 × 101 grid over output and understanding values) and the paper's evolutionary conclusions depend on the rare-exploration assumption it studies.
• No empirical hiring data are analyzed; the equilibria are analytic and the claims about which one markets occupy are inferential.
• The argument is aimed at mathematics and theoretical sciences, where output is relatively easy to fake with AI, and the authors note that the mechanism is weaker in empirical fields.

Citation

Pal, S., & Park, P. S. (2026). Game theory of academic effort and hiring in the age of AI. arXiv:2610.35802.