🔬 physics

The Evolutionary Dynamics of AI, Politicization, Contestation, and Trust in Science Funding

This paper employs an evolutionary game-theoretic model to demonstrate that perceived political interference in science funding can trigger cascading resistance among scientists, leading to divergent institutional outcomes where operational capacity and public legitimacy are governed by separate, often decoupled, conditions heavily influenced by the public's acceptance of AI-assisted review systems.

Original authors: Animesh Ray

Published 2026-07-30

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Animesh Ray

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a giant, high-stakes game of "Show and Tell" that runs the engine of our modern world. In this game, brilliant scientists come up with wild ideas for new medicines, cleaner energy, and faster computers. To get the money to build these ideas, they have to convince a panel of experts that their plans are good. This process is called peer review, and it's the gatekeeper that decides which projects get funded and which ones go home empty-handed. For this system to work, two things must happen: first, the experts must actually do the work of reading the proposals (capacity), and second, the rest of us—the public—must believe the process is fair and not rigged by politicians (legitimacy). If the public thinks the game is unfair, they stop trusting the results, and the whole system can crumble, even if the scientists are still doing their jobs.

Now, imagine the government decides to speed things up by bringing in a super-fast robot judge (Artificial Intelligence) to help read the proposals, especially if the human experts start complaining that the system is unfair. This is where a new study by Animesh Ray steps in. The author built a complex computer simulation—a digital sandbox—to see what happens when scientists, the public, and the funding agency all start reacting to each other in a chaotic, politicized environment. The study suggests that simply swapping humans for robots doesn't fix the problem; in fact, it might create a "zombie" version of the funding agency that keeps churning out decisions but has lost all its soul and public trust.

The Digital Sandbox: A Game of Strategy

In this simulation, the author treats the funding world like a three-way tug-of-war involving Scientists, the Public, and the Funding Agency.

The Scientists have a few moves. They can play nice, review their colleagues' work, and submit their own proposals. But if they feel the system is rigged or unfair, they can start a protest. They might refuse to review anyone's work, or they might try to "flood" the system by submitting a massive number of proposals just to clog the pipes. It's like a group of kids in a game deciding to either play by the rules or, in a fit of pique, throw a thousand balls at the referee at once.

The Public is watching closely. They want to know if the game is fair. If they see the agency relying too much on the robot judge (AI) without a good explanation, or if they see scientists fighting, they might stop trusting the whole thing. They can accept the new system, reject it, or even try to sue (litigate) to stop it.

The Funding Agency is the referee trying to keep the game going. It has a limited budget and a growing pile of proposals. When the scientists start protesting, the agency tries to fix it by doing two things: paying the human reviewers more money to keep them happy, and turning up the volume on the AI robot judge to process the backlog faster.

The Big Surprise: The "Zombie" Agency

The simulation revealed some startling and counterintuitive results. The most important finding is that operational capacity (how many proposals get processed) and institutional legitimacy (how much the public trusts the system) are two separate things that can fail independently.

The study suggests that an agency can successfully use AI to clear a massive backlog of proposals, keeping the money flowing and the robots working at full speed. However, at the exact same time, the public can completely lose faith in the system, viewing the decisions as illegitimate. The author calls this a "zombie" state: the institution is technically alive and moving, but its authority is dead. It's like a restaurant that serves food incredibly fast because a robot is cooking, but the customers hate the food and the chef, so they stop coming in, even though the kitchen is running at 100% capacity.

The "Proposal Flood" and the Budget Trap

One of the most vivid scenarios the paper explores is what happens when scientists get angry. Instead of just refusing to work, they might decide to "flood" the system with thousands of extra proposals. The simulation shows that this strategy is incredibly effective at breaking the system. Even if the AI robot is super-fast, it can't keep up if the scientists keep throwing more and more proposals at it.

Here's the twist: the agency tries to fix this by paying the human reviewers more money to get them back on board. But the study suggests this can backfire. If the public sees the agency handing out huge "emergency" payments to scientists during a crisis, they might view it as a bribe. This perception can destroy whatever trust was left, making the situation worse. Furthermore, paying the humans more drains the budget, which means there's less money left to pay for the AI robots, creating a vicious cycle where the agency can't afford to fix the problem it's trying to solve.

The Role of Chance and Timing

The paper also highlights that the outcome isn't always predictable. In the simulation, identical starting conditions could lead to totally different endings just because of small, random differences in how people reacted early on. Sometimes, a little bit of extra trust at the very beginning could save the system; other times, a tiny spark of distrust could cause a total collapse. The study suggests that once the system tips into a "zombie" state or a total collapse, it's very hard to pull it back.

What the Study Does Not Say

It is important to note what this paper does not claim. It does not say that AI is bad or that we should never use it. The simulation actually shows that AI can process proposals effectively. The problem isn't the technology itself; it's the trust surrounding it. The study also doesn't say that scientists are inherently bad or that they will definitely flood the system; rather, it suggests that if the political environment gets bad enough, the incentive to flood the system becomes a rational choice for them.

The Bottom Line

In simple terms, this study suggests that you can't just buy your way out of a trust crisis with faster robots or bigger paychecks. If the public thinks the game is rigged, no amount of AI efficiency will fix it. The agency might keep the lights on and the proposals moving, but if the people watching lose faith, the whole enterprise becomes a hollow shell. The key takeaway is that fixing the "trust" part of the equation is just as important as fixing the "speed" part, and they are two different problems that need two different solutions. If we don't pay attention to the human feelings and the perception of fairness, we risk building a scientific funding system that works perfectly on paper but fails in the real world.

Technical Summary: The Evolutionary Dynamics of AI, Politicization, Contestation, and Trust in Science Funding

Problem Statement
Modern democratic societies rely on public funding for scientific research, a system predicated on the dual pillars of operational capacity (processing proposals) and institutional legitimacy (public trust in fairness). This system faces a dual threat: the perception of political interference in funding decisions and the potential substitution of human peer reviewers with Artificial Intelligence (AI). The paper addresses a critical gap in understanding how these factors interact. Specifically, it investigates whether the operational efficiency gained by adopting AI can compensate for the erosion of institutional legitimacy caused by perceived political bias and the displacement of human expertise. The central question is whether capacity and legitimacy are coupled or independent dynamics, and how they evolve when scientists resist the system (via refusal to review or proposal flooding) and the public reacts to AI adoption.

Methodology
The authors employ an evolutionary game-theoretic model to simulate the co-evolutionary dynamics of three interacting populations:

Scientists: A population choosing among four strategies: normal participation ( $N$ ), refusal to review ( $R$ ), proposal intensification/flooding ( $P$ ), or a combination of both ( $B$ ).
The Public: A population choosing to accept AI-assisted review ( $A$ ), publicly reject it ( $J$ ), or litigate ( $S$ ).
The Funding Agency: Modeled not as an evolving strategic population but as an adaptive controller. It adjusts two policy levers in response to the other populations: the level of AI adoption ( $\phi$ ) and the financial remuneration for human reviewers ( $\rho$ ), constrained by a finite budget.

The system is driven by pairwise-comparison imitation dynamics (Fermi dynamics), where agents adopt strategies based on relative payoffs. The Agency acts as a capacity-responsive controller, increasing AI deployment to close the gap between proposal load and human review capacity, while adjusting reviewer pay to incentivize participation. The model distinguishes between operational capacity (ability to process proposals) and institutional legitimacy (public trust), treating them as separate state variables that can fail independently.

Simulations were conducted using deterministic trajectories and finite-population stochastic processes (Wright–Fisher resampling and individual-level pairwise comparison). The study utilized parameter sweeps, Latin-hypercube sampling, and principal component analysis (PCA) to map the regime landscape and identify attractors.

Key Contributions and Results

Decoupling of Capacity and Legitimacy:
The model demonstrates that operational capacity and institutional legitimacy obey separate dynamics and can fail independently.
- Capacity Failure: Even with maximal AI deployment, the system can fail to clear backlogs if scientists engage in "proposal intensification" (flooding the system with proposals while refusing to review). The binding constraint is often the shared budget: resistance drives up reviewer pay, which drains the budget, thereby suppressing the Agency's ability to fund AI operations, regardless of technical AI capacity.
- Legitimacy Failure: Conversely, the system can achieve full operational recovery (clearing backlogs) while public trust collapses completely, resulting in a "zombie" institution that functions technically but lacks authority.
The "Proposal Flooding" Mechanism:
Contrary to the intuition that scientists would primarily resist by refusing to review, the model finds that proposal intensification is the dominant retaliatory strategy. Scientists may continue to review (avoiding direct career sanctions) but flood the agency with proposals. This strategy overwhelms the system faster than AI can scale, leading to persistent congestion even when human refusal rates are low.
The "Bribe-Signal" Effect of Emergency Pay:
The model identifies a counterintuitive dynamic where increasing reviewer pay to secure cooperation during a crisis can backfire. If the public perceives emergency pay as a "bribe" to quell resistance, it accelerates the erosion of institutional legitimacy. This effect is active primarily in high-resistance regimes where pay remains elevated; in low-resistance regimes, pay returns to baseline, rendering the bribe-signal inactive.
Bistability and Sensitivity to Initial Conditions:
The system exhibits bistability regarding legitimacy. The final state (recovery vs. collapse) is determined not by the speed of institutional repair or the extent of AI adoption, but by the public's acceptance dynamics and initial conditions (e.g., initial resistance seeds).
- Stochastic Outcomes: In finite populations, identical institutions can reach markedly different fates (e.g., joint survival, joint failure, or legitimacy recovery without capacity) based on stochastic drift.
- Algorithm Dependence: The frequency of specific outcomes (such as the "zombie" state) depends heavily on the assumed noise process. Under a Wright–Fisher resampling layer, "zombie" states occur frequently; under an individual-level pairwise-comparison (Fermi) process, they are absent, with the system bifurcating only between "legitimacy recovery without capacity" and "joint failure."
Thresholds and Phase Transitions:
- Resistance Threshold: Resistance cascades ignite when politicization levels exceed a critical threshold relative to the career cost of refusal.
- Legitimacy Threshold: AI-review legitimacy follows a sharp binary transition. It survives only if the rate of public acceptance gain exceeds the combined erosion from rejection, litigation, and resistance.
- Capacity Threshold: Full operational recovery requires the product of AI productivity and the deployment ceiling to exceed the proposal-load multiplier ( $\beta_{AI} \cdot \phi_{max} > \chi$ ).

Significance and Claims
The paper claims to provide a rigorous, quantitative framework for understanding the structural vulnerabilities of science funding systems in the age of AI and political polarization. Its primary significance lies in identifying that automation does not equal trust. The model suggests that policymakers cannot rely on AI to solve capacity crises without simultaneously addressing the legitimacy crisis, as the two are governed by distinct mechanisms.

The work highlights that:

Operational fixes (AI) do not solve legitimacy problems: Restoring throughput does not automatically restore public trust.
Legitimacy fixes do not solve capacity problems: Increasing public acceptance can restore trust even while the system remains operationally clogged.
Early intervention is critical: Because the system is sensitive to initial conditions and stochastic drift, early actions that shift public acceptance or scientist resistance can determine the long-term fate of the institution.
Policy Levers: The most effective levers for policy are not necessarily the speed of AI adoption or the level of reviewer pay, but rather the management of politicization, the reduction of career costs for resistance, and the design of transparent AI rollouts that maximize public acceptance gains.

The authors conclude that the "AI revolution" presents a stress test for the social contract of science funding. Without a renegotiation of trust mechanisms, the substitution of human expertise with automated systems risks unraveling the institutional legitimacy of the entire funding enterprise, potentially leading to a state where the pipeline continues to function while the public authority of science collapses. The paper explicitly calls for future empirical calibration of its parameters (e.g., price sensitivity of experts, trust recovery rates) to refine these theoretical predictions.