How People Really Cooperate: A New Method Reveals the Hidden Rules of Collective Action

In the xAire experiment, 77 groups of six strangers played a climate cooperation game in public spaces across Barcelona. Just 39% of them avoided collective failure. But within that stark outcome lies a deeper truth: every decision to contribute—or hold back—was shaped by at least four competing impulses. Now, a new data-driven method has cracked open the black box of human behavior in these high-stakes dilemmas, revealing not one dominant strategy, but an ensemble of behavioral heuristics, each pulling in different directions. When simulated using a simple plurality vote—where the most common suggestion wins—these competing rules reproduce real-world cooperation patterns with striking fidelity, outperforming both single-rule models and random chance. This isn’t just a better simulation. It’s a new way to understand how people actually make decisions under pressure, with implications for climate policy, public health campaigns, and any system that depends on collective action.

The Science

Collective-risk dilemmas (CRDs) are the laboratory version of real-world crises like climate change: individuals must contribute personal resources to a shared goal, or everyone suffers a common loss. The classic setup, pioneered by Milinski et al. (2008), pits self-interest against group survival. If the group meets its target—say, raising enough money to avoid a simulated environmental disaster—everyone keeps what’s left of their endowment. If not, all lose everything. These experiments generate rich data: who contributed what, when, and under what conditions. But until now, interpreting that data has meant choosing between oversimplified models or opaque machine learning algorithms that don’t reveal how people decide.

The method introduced by Sadurní, Díaz, Perelló, and Montero (2026) flips this approach. Instead of assuming people follow a single strategy—like always copying others or always contributing half their budget—their framework treats each person as a committee of behavioral heuristics, each responding to a different piece of information. These aren’t abstract rules; they’re derived directly from real human behavior in the xAire experiment, a citizen science project that turned public plazas in Barcelona into pop-up behavioral labs. Participants played a 10-round game where they could contribute 0, 2, or 4 monetary units (MU) per round, with a collective target of 120 MU—half the total initial endowment.

The researchers identified four key feedback channels visible to players after each round, and built one heuristic for each:

• Hself: The tendency to repeat one’s own past action. If you gave 2 MU last round, how likely are you to give 2 again?
• Hpeer: The response to peer behavior. If others gave more on average, do you give more too?
• Hfund: Sensitivity to collective progress. If the group is far from the target, do you contribute more to catch up?
• Hbudget: Awareness of personal resource depletion. If you’ve already spent 60% of your starting funds, do you tighten your belt?

Each heuristic is a conditional probability table—essentially a lookup chart—mapping a specific game state (e.g., “others gave 3 MU on average”) to the likelihood of contributing 0, 2, or 4 MU. These tables are built entirely from empirical data, with no assumptions about rationality or optimization. Even the first move is seeded from real first-round behavior (see Table 1).

The innovation lies in how these heuristics are combined. In each round, every simulated agent evaluates all four heuristics, producing a decision vector—say, (2, 4, 2, 0)—where each number is the contribution suggested by one rule. The final action is determined by a plurality vote: the most frequent suggestion wins. If there’s a tie—like two 2s and two 4s—one is chosen at random. This hard-voting rule adds no free parameters; it’s a democratic arbitration among behavioral impulses. The entire process is non-parametric, reproducible, and grounded in observed behavior.

What They Found

The real test wasn’t whether the model could run—it was whether it could replicate the messy, uneven patterns of real human groups. To find out, the researchers simulated 77 games (matching the original dataset size) and compared synthetic outcomes to empirical ones across five macro-indicators, all restricted to successful games (those that met the target). The results were striking.

First, the ensemble model outperformed both single-heuristic baselines and a uniform-random benchmark. No single heuristic—whether self-consistency, peer mimicry, progress tracking, or budget awareness—could reproduce the full range of observed behaviors. But the ensemble, using the same data, came remarkably close.

Take the final payoff distribution. In real games, players kept varying amounts of their endowment, creating a lopsided distribution shaped by inequality in starting wealth and contribution patterns. The ensemble model reproduced this with high fidelity, as measured by the Jensen–Shannon divergence and Wasserstein-1 distance—two statistical metrics that quantify how similar two distributions are. For players who started with 60 MU, the synthetic and empirical payoff distributions were nearly indistinguishable (JS divergence: 0.012; Wasserstein distance: 0.048).

Similarly, the model captured inequality in final payoffs across treatments. In the Unequal-H treatment—where four players started with 30 MU and two with 60 MU—real groups ended with a Gini coefficient of 0.18, indicating moderate inequality. The ensemble simulation produced a Gini of 0.17, a difference of just ΔG = 0.01. The area between the empirical and synthetic Lorenz curves—the standard way to visualize inequality—was just 0.025, confirming that the model preserved the social structure of outcomes.

Average Contribution per Round by Starting Endowment

Label	Value
Empirical (30 MU)	2.1
Synthetic (30 MU)	2
Empirical (60 MU)	2.8
Synthetic (60 MU)	2.7
Empirical (48 MU)	2.4
Synthetic (48 MU)	2.3
Empirical (24 MU)	1.9
Synthetic (24 MU)	1.8

Source: xAire experiment data and ensemble simulation (Sadurní et al., 2026). Lines show average contributions at two time points.

Even more impressive was the model’s ability to reproduce dynamic patterns. The average contribution per round, broken down by starting endowment, followed a characteristic arc: high early contributions, a dip in the middle, and a surge near the end. The ensemble simulation matched this trajectory almost perfectly, with a normalized mean-squared error (nMSE) of just 0.034 and a dynamic time warping (DTW) score of 0.11—both significantly better than any single heuristic.

The same was true for the common fund trajectory: the cumulative amount contributed by the group over time. Real groups tended to lag early, then accelerate as the deadline approached. The ensemble model replicated this ramp-up effect, while single-heuristic models either over-contributed early (Hfund) or under-contributed throughout (Hself).

Final Payoff Distribution: Empirical vs. Synthetic

Label	Value
0-10 MU	0.02
10-20 MU	0.05
20-30 MU	0.1
30-40 MU	0.18
40-50 MU	0.25
50-60 MU	0.4

Source: xAire experiment and ensemble simulation (Sadurní et al., 2026). Bins of 10 MU. JS divergence: 0.012.

But the most revealing result came from the decision vectors themselves. In 41% of all decisions, the four heuristics disagreed—they proposed different contributions. Yet in 59% of cases, at least two heuristics aligned, and a clear plurality emerged. When conflict arose, the random tie-breaker ensured behavioral diversity without introducing bias. This suggests that human decision-making in CRDs isn’t monolithic; it’s a constant negotiation between competing impulses, with no single rule dominating.

Why This Changes Things

For decades, behavioral science has treated cooperation as a puzzle to be solved with incentives, nudges, or appeals to morality. Game theory offers elegant models—like tit-for-tat or conditional cooperation—but they often fail to predict real-world outcomes. Machine learning models can predict behavior, but they’re black boxes: we don’t know why they work.

This new method bridges that gap. It doesn’t assume people are rational optimizers. It doesn’t reduce them to a single behavioral type. Instead, it treats behavior as an emergent property of multiple, coexisting heuristics—simple rules of thumb that respond to different cues in the environment. And because those heuristics are derived from data, not theory, the model is both grounded and flexible.

The implications extend far beyond the lab. Consider climate policy. Current models often assume nations will act based on cost-benefit analyses or peer pressure. But what if countries are also responding to their own past commitments (Hself), the pace of global progress (Hfund), and domestic budget constraints (Hbudget)? A model that captures all four could generate more realistic scenarios for international negotiations.

Or take public health. During a pandemic, individuals decide whether to mask, vaccinate, or isolate based on a mix of personal experience, community behavior, case counts, and personal risk. A heuristic ensemble could simulate how different messaging—emphasizing peer behavior vs. collective progress—might shift the balance of these impulses.

Even within organizations, this framework could help design better teamwork systems. If employees are torn between sticking to their routine (Hself), following team norms (Hpeer), hitting quarterly targets (Hfund), and managing workload (Hbudget), leaders could use this model to identify which levers matter most.

The method also challenges a deep assumption in behavioral science: that people have stable, identifiable types—“cooperators,” “free-riders,” “reciprocators.” The xAire data suggests otherwise. People aren’t one thing. They’re a shifting coalition of impulses, and their behavior emerges from which impulse wins the vote in a given moment. This doesn’t mean behavior is unpredictable. It means it’s context-dependent in a structured way.

And because the model is data-driven and modular, it can be adapted to any decision-making context with observable feedback. Want to study charitable giving? Plug in heuristics for past donations, peer giving, campaign progress, and personal income. Studying energy conservation? Use past usage, neighborhood averages, grid demand, and household budget. The structure stays the same; only the heuristics change.

What's Next

The researchers acknowledge several limitations. The current model uses first-order Markovian heuristics—each decision depends only on the previous round. But people might have longer memories, or respond to trends over time. Future versions could incorporate multi-round patterns or learning within games.

The plurality vote is simple, but other arbitration rules might work better. A weighted vote—where some heuristics dominate based on past accuracy—could improve fit, though at the cost of adding parameters. Or a sequential override system—“if Hfund says ‘give 4,’ do it, otherwise check Hpeer”—might better reflect how people prioritize concerns.

The model also doesn’t explain why certain heuristics dominate in certain contexts. Why does Hself have such strong predictive power? Is it habit, inertia, or a cognitive shortcut? These questions require deeper psychological investigation.

And while the model reproduces observed patterns, it hasn’t yet been tested in unseen scenarios. The real power of such a framework lies in its ability to simulate interventions: What if the target were higher? What if inequality were worse? What if feedback were delayed? The authors note that synthetic data generation is a key feature, but full exploration of counterfactuals remains future work.

Still, the core insight stands: cooperation isn’t driven by a single force. It’s a tug-of-war between self, peers, progress, and resources. And the winner isn’t predetermined—it’s decided, round by round, by a quiet internal vote.

That changes how we think about collective action. We don’t need to turn everyone into altruists. We just need to design systems where the right heuristics win the vote more often. If we can amplify the voice of Hfund—the one that cares about the common goal—while quieting the one that says “I’ve done my part,” we might just meet our targets, one plurality at a time.