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

• Publications serve as both a scientific record and, historically, evidence of expertise that shapes early-career job prospects.Source: §1 · p. 2
• In mathematics and theoretical sciences, accessible AI weakens the once-tight link between strong research results and field understanding.Source: §1 · p. 2
• The paper models applicants’ study choices and recruiters’ hiring decisions as strategically interdependent.Source: §1 · p. 3
• It characterizes all pure-strategy perfect Bayesian equilibria and examines which behaviors emerge through payoff-based imitation and rare exploration.Source: §1 · p. 3
• The analysis relates institutional reforms to model-parameter changes, yielding conditional predictions about their effects on study and hiring.Source: §1 · p. 3

• Applicant profiles encode few or many papers, and the model assumes candidates cannot fake their resumes.Source: §2 · p. 4
• The test always passes candidates with full understanding, while candidates without it fail with probability τ and may pass as false positives.Source: §2 · p. 4
• Deep study guarantees full understanding and costs more than shallow study, with the extra cost higher for medium-potential than high-potential applicants.Source: §2 · pp. 4, 6
• High-potential applicants produce many papers under either trajectory, while medium-potential applicants do so only with shallow + AI; low-potential applicants produce no papers and have no understanding.Source: §2 · p. 5
• Recruiters observe paper output and pass/fail results, then choose whether to accept based on their separate valuations of papers and understanding or their opting-out payoff.Source: §2 · pp. 4, 6

The five pure-strategy equilibria of the baseline model

• The baseline model has five equilibrium patterns: E1–E4 can coexist when A<r, while E5 is the unique equilibrium when A≥r.In E5, all applicants with papers are hired and nobody invests in understanding because output alone makes hiring worthwhile.Source: The five pure-strategy equilibria of… · pp. 8–9
• E1 is a self-sustaining trap: everyone studies shallowly, the test reveals nothing, and the recruiter rejects everyone.A deep-work deviation bears a cost but cannot be distinguished from a lucky AI user.Source: The five pure-strategy equilibria of… · pp. 8–9
• In E2, only high-potential applicants study deeply, and many-paper passers are hired when B≥B1 and τW≥δH.The passed test must screen enough AI users for the job premium to cover high-potential applicants’ extra study cost.Source: The five pure-strategy equilibria of… · pp. 8–9
• In E3, all strategic applicants study deeply and passers are hired; in E4, only medium-potential applicants study deeply and few-paper passers are hired.E3 requires hiring few-paper passers; E4 leaves high-potential applicants shallow and unemployed because their output cannot be hidden.Source: The five pure-strategy equilibria of… · pp. 8–9

Extra equilibria in the extended super-output model

• With super-output, two new equilibria arise for r/2≤A<r: high-potential applicants stay shallow, while medium-potential applicants either stay shallow or study deeply.In E6, only high-potential applicants are hired for super-output; in E7, medium-potential applicants can also be hired through the few-paper channel.Source: Extra equilibria in the extended… · p. 12
• Super-output lowers the output-valuation threshold for high-potential applicants to abandon understanding from A=r to A=r/2.When A≥r, the baseline output-dominant regime returns: applicants with at least one paper are hired regardless of the test, and nobody studies deeply.Source: Extra equilibria in the extended… · p. 12
• The share of hired candidates who understand is highest in E2 or E3; which performs better depends on applicant-type distribution, not test accuracy.E2 makes errors when medium-potential applicants pass; E3 makes errors when low-potential applicants pass, and τ scales both error rates equally.Source: Extra equilibria in the extended… · pp. 12–13
• D1 can eliminate E4 when high-potential applicants would gain from deviating, but E7 survives because its off-path response does not change hiring or study choices.In E7, D1 restricts the response when δM>τW; when δM≤τW, either response is consistent with D1.Source: Extra equilibria in the extended… · p. 15

• The simulations track high- and medium-potential applicants and recruiters as they imitate more successful strategies, with rare exploration and learning confined to each population.The process estimates long-run strategy frequencies rather than assuming that the populations coordinate on a particular equilibrium.Source: §4 · pp. 15–16
• When A≥r, both strategic applicant types learn shallow study, and recruiters learn to hire every many-paper applicant regardless of test outcome.With an accurate test, recruiters also occasionally hire few-paper applicants who pass; these are low-potential applicants who pass by luck, and accepting them is costly but only weakly selected against.Source: §4 · p. 16
• For A<r, sufficiently high B and accurate testing support depth; when B1<B<B0, high-potential applicants go deep while recruiters hire many-paper applicants who pass.For B>B0, recruiters also hire few-paper applicants who pass, rewarding depth by medium-potential applicants; their depth dominates at low A.Source: §4 · p. 16
• At test accuracy τ=0.5, high-potential applicants rarely go deep, while medium-potential applicants do so only when B is very high and A is low.Source: §4 · p. 16
• A=r/2 replaces A=r as the threshold beyond which depth disappears when high-potential applicants can produce super-output through shallow study.At test accuracy τ=0.5, neither applicant type goes deep anywhere; at higher accuracy, high-potential applicants study deeply for A<r/2 and high B, but less often than in the baseline.Source: §4 · pp. 16, 19, Figure 4 · p. 18
• Super-output can unravel depth even where it remains an equilibrium, because recruiters may drift into accepting it when no applicant produces it.When simulations bar recruiters from accepting super-output, high-potential applicants’ depth rises above its baseline level.Source: §4 · p. 19

5 Discussion

• Deep-study equilibria coexist with an all-shallow equilibrium under the same fundamentals, with expectations determining whether a pass signals understanding or luck.Recruiters’ expectations shape how they interpret passes, while applicants’ expectations shape whether deep study seems worthwhile.Source: §5 · p. 20
• Escaping the shallow-study trap requires departments to commit together to rewarding understanding, but rare-exploration learning often penalizes either side for changing alone.Faster exploration may ease coordination; the trap is especially persistent when tests are less accurate or AI amplifies high-potential output.Source: §5 · p. 20
• In reverse-sorting equilibrium E4, medium-potential applicants study deeply while high-potential applicants rely on AI, because high-potential output makes them appear shallow.Few papers plus a pass signals deep study for medium-potential applicants, while many papers prompt rejection even when high-potential applicants study deeply.Source: §5 · pp. 20–21
• High-potential applicants study deeply only when they believe the test is accurate enough, so transparent standards matter when applicants might underestimate accuracy.A rigorous test applicants believe is easy to pass does not deter shallow study.Source: §5 · p. 21
• More accurate tests alone may not encourage understanding if output alone warrants hiring; hiring every passer raises hiring rates but can admit low-potential applicants who passed by luck.Hiring only many-paper applicants who pass screens out low-potential applicants but admits some medium-potential AI users; which rule yields more understanding per hire depends on applicant-pool composition, not test accuracy.Source: §5 · p. 21
• Oral defenses and chalk talks raise test accuracy, while publication caps lower the weight placed on output.The model maps these reforms to changes in test accuracy and recruiter valuation of output, respectively.Source: §5 · p. 22

5.1 Related work

• The paper combines static game theory with evolutionary dynamics and adds applicant understanding as a second variable alongside output in job-market signaling.It applies evolutionary dynamics to study which outcomes emerge through learning.Source: §5.1 · pp. 22–23

5.2 Limitations

• The model assumes binary understanding, deterministic output, externally fixed study costs, applicant composition and test accuracy, and no understanding gain from AI-assisted research.These assumptions exclude uncertain learning, variable output, feedback between behavior and model inputs, and AI interactions that build deep understanding.Source: §5.2 · p. 23

Imitation and rare exploration in applicant and recruiter populations

• Each of three finite populations updates separately; high- and medium-potential applicants choose deep or shallow study, recruiters choose pure acceptance rules, and low-potential applicants remain fixed.Recruiters additionally condition on super-output in the extension.Source: Imitation and rare exploration in… · p. 28
• Payoffs are expected against other populations’ current strategies; applicant hiring chances average over tests and recruiter rules, while recruiters face exogenously set applicant-type proportions.No agent’s payoff depends on the composition of its own population.Source: Imitation and rare exploration in… · p. 28
• At each step, one population is selected; focal agents usually imitate same-population role models by the Fermi rule, but explore another available strategy with probability ε.The imitation probability depends on the payoff difference between role model and focal agent.Source: Imitation and rare exploration in… · p. 29
• As ε→0, the process spends nearly all its time in monomorphic states; strategy introductions generate transitions via takeover probabilities, yielding a unique stationary distribution solved exactly.The baseline has 64 such states and the super-output model has 128.Source: Imitation and rare exploration in… · p. 29
• The analysis reports long-run action frequencies for applicant populations and recruiter acceptance frequencies after observations; payoff-neutral switches occur after unproduced observations, while rare observations are weakly selected.Two recruiter rules differing only after an observation no current applicant produces earn identical payoffs.Source: Imitation and rare exploration in… · p. 29