AI can keep grant pipelines running even as public trust evaporates, producing “zombie” funding agencies; small early resistance and signaling choices (for example emergency pay) can flip identical institutions between recovery and collapse.

Main Finding

A co-evolutionary evolutionary-game model shows that automation (AI-assisted peer review) can preserve operational processing capacity while institutional legitimacy collapses — creating “zombie” funding agencies that keep paying out grants but have lost public trust. Capacity and legitimacy follow distinct dynamics and can fail independently; outcomes are highly sensitive to initial resistance, politicization, and stochastic finite-population effects. Policy levers that only raise throughput (more AI, higher pay) can backfire on legitimacy unless trust is proactively managed.

Key Points

• Model actors and strategies

  • Three interacting actors: Scientists (population), the Public (population of active non-scientists), and an adaptive Funding Agency (controller).
  • Scientists choose among: normal participation (N), refusal to review (R), proposal intensification/flooding (P), or both refusal+flooding (B).
  • Public chooses: accept AI review (A), publicly reject (J), or litigate (S).
  • Agency adapts AI deployment (φ ∈ [0,1]) and reviewer pay (ρ ∈ [0,1]) subject to a finite budget.

• Five principal theoretical findings

• Capacity vs legitimacy decouple: AI can close backlogs without restoring public trust; legitimacy can recover even when capacity remains constrained.
• Legitimacy is bistable: public acceptance often determines long-run state; small differences push the system to distinct basins of attraction.
• Low individual cost to refuse reviewing enables resistance cascades: proposal flooding can outpace AI adoption and collapse capacity.
• Emergency reviewer pay is double-edged: raising pay can stabilize participation only within strict budget/solvency bounds; above these bounds pay can be perceived as a bribe and erode public trust.

• Stochastic sensitivity: in finite populations identical institutions may reach very different outcomes by chance; how often chance favors recovery depends on the stochastic update rule (noise model), not on a single fixed failure probability.

• Important emergent regimes identified

  • Joint survival (capacity and legitimacy survive).
  • Capacity-only (technical throughput preserved, legitimacy lost — “zombie” agency).
  • Legitimacy-without-capacity (public trust returns but backlog persists).
  • Joint failure (both capacity and legitimacy collapse).
  • Persistent oscillations/regimes near thresholds under certain pay incentives.

• Policy warnings

  • Automation ≠ trust: scaling AI without addressing legitimacy risks permanent public rejection.
  • Prepare for strategic retaliation (proposal flooding), which can overwhelm AI if not anticipated.
  • Transparent, explainable AI rollouts and legitimacy-focused interventions are urgent priorities; reactive emergency pay can worsen trust problems.

Data & Methods

• Modeling framework

  • Analytical and numerical evolutionary-game model (finite populations) with co-evolution among Scientist strategies, Public strategies, and an adaptive Agency controller.
  • Scientists and Public update strategies via pairwise-comparison (Fermi) imitation dynamics; weak-selection approximations yield replicator-like forms for analytic insight.
  • Agency computes required AI deployment φ_req = clip[(r + χ p)/β_AI, 0, φ_max] (r = fraction refusing review, p = fraction intensifying submissions, χ = submission multiplier, β_AI = reviews/unit AI) and adjusts actual φ toward φ_req subject to fiscal constraints and litigation effects; reviewer pay ρ responds to resistance separately.

• Payoffs

  • Payoff matrices specified for Scientists and Public that include career benefits, reviewer remuneration, collective-action benefits, conformity gains, politicization parameter G, institutional legitimacy L, AI-review legitimacy L_A, costs for litigation, and submission/competition effects.
  • Reviewer pay increases can enter Public payoff as a negative “bribe” signal.

• Simulations

  • Deterministic adaptive trajectories integrated using full Fermi dynamics; numerical regime diagnostics at checkpoints (2k, 5k, etc.), runs up to 50k time-steps for principal scenarios.
  • Fixed-point and cycle detection based on amplitudes and derivatives over windows; parameter sweeps explored stability/threshold structure.
  • Finite-population stochastic simulations also run to probe drift across basins; outcomes depend on the stochastic update/noise model.

• Analytical results

  • Derived threshold conditions: when refusal/intensification exceed critical levels, resistance becomes self-reinforcing; political G lowers those thresholds; legitimacy survives when repair rates (by acceptance/participation) exceed erosive forces (politicization, perceived bribe effects).
  • Showed formal separation of capacity condition (depends on r,p, technical ceiling) from legitimacy dynamics (depends on L, L_A, public perceptions).

• Limitations noted by authors

  • Abstract/theoretical: parameters (e.g., politicization G, costs of refusing review, AI legitimacy signals) require empirical calibration.
  • Litigation and legal costs modeled via public strategy payoffs but not fully endogenous to agency budgeting.
  • Simplifying assumptions about bounded strategy sets and agent homogeneity.

Implications for AI Economics

• Efficiency vs legitimacy trade-offs: Investments that increase throughput (AI tools) generate private and social returns only if institutional legitimacy remains intact. Economic assessments of AI adoption in public institutions must include legitimacy externalities and political economy feedbacks, not only per-unit productivity gains.
• Endogenous contestation and rent-seeking: Researchers’ strategic behavior (flooding, boycotts) is a rent-seeking/collective-action problem that can negate AI capacity gains. Models of R&D funding and innovation policy should internalize strategic submission incentives and enforcement/monitoring costs.
• Budget-constrained adoption decisions: The Agency’s optimization of AI deployment and reviewer pay under budget constraints highlights non-linearities — marginal increases in pay may produce discontinuous legitimacy effects; welfare analysis needs to consider non-monotonic responses.
• Value of transparency and explainability: From an economic-design perspective, credible signaling (transparent, explainable AI systems, third-party audits) can shift basins of attraction toward legitimacy-preserving equilibria. Cost–benefit analyses should incorporate costs of transparency measures as investments in institutional capital.
• Stochastic fragility and policy timing: Finite-population stochastic effects imply path-dependence and early-intervention value. Economists should account for the option value of early legitimacy-preserving actions and the potentially high social cost of delayed responses.
• Regulatory and legal externalities: Litigation dynamics can rollback AI deployment and impose unmodeled fiscal/legal costs; regulatory design (rules about AI use in peer review, disclosure requirements) materially changes equilibria and thus the social returns to AI adoption.
• Empirical priorities for AI economics research: Estimate parameters central to dynamics — cost to careers from refusing reviews, typical submission elasticity (χ), public responsiveness to AI-legitimacy signals, litigation probabilities and costs, and β_AI (reviews per AI unit). These calibrations are essential to quantify welfare impacts and to design robust rollout strategies.

Suggested actionable priorities for policymakers and economists - Prioritize trust-building: invest in explainability, independent audits, and participatory rollout designs before scaling automated review. - Model and simulate strategic submission behavior to size capacity needs under worst-case flooding scenarios. - Avoid ad-hoc “emergency pay” without public signaling strategies that frame payments as fair compensation rather than special favors. - Gather empirical estimates for key model parameters to move from qualitative warnings to quantitative policy prescriptions.