When community beliefs steer AI toward cooperation—or free-riding

A community’s support for cooperation can be read in two ways: as a norm to follow, or as an opportunity to benefit from someone else’s effort. A new preprint reports that explicitly describing which considerations matter can shift language models between these responses.

The study, “Exploit or conform? How norm utility prompting changes the way LLMs use community beliefs,” by Leon Klingborg, Theodor Cimpeanu and Alexander J. Stewart, was posted on November 2, 2026, as arXiv:2611.04048v1. Its findings concern model-generated choices in simplified scenarios, not demonstrated improvements in real-world cooperation.

What the models were asked to decide

The researchers tested three open-weight reasoning models: Qwen 3.8 27B, Gemma 4 31B and Muse Glimmer 30B. Each faced a one-shot, two-player decision, with choices neutrally labeled “action 1” and “action 2” rather than “cooperate” and “defect.”

In the setup, cooperation produces a benefit for both players. A lone cooperator pays the entire cost; two cooperators split it. If neither cooperates, neither receives anything. The researchers used two payoff settings: Snowdrift, where cooperating is financially worthwhile if the other player does nothing, and the Prisoner’s Dilemma, where defection gives the individual a higher material payoff regardless of the partner’s choice.

Prompts specified what percentage of a wider community believed cooperation was good. That percentage changed from zero to 100 in five-percentage-point steps. For each model and experimental setting, the researchers estimated cooperation rates from 1,000 generated responses, focusing on reasoning-enabled runs at temperature 1, their maximum sampling temperature.

Four main conditions crossed two changes: whether models had to publicly state what others ought to do, and whether prompts explicitly introduced “norm utility.” Here, that meant weighing material gain alongside conformity to community beliefs and consistency between beliefs and actions. The prompts left models to decide those weights.

The same information, opposite trends

In Snowdrift, cooperation generally declined as community support for cooperation increased when norm utility was not explicitly prompted. The authors interpret this as models treating community beliefs as information about a partner’s likely behavior: if someone else will probably contribute, free-riding becomes attractive.

That interpretation is plausible, but community beliefs were not actual measurements of the partner’s behavior. Nor were the response curves uniformly downward: under the simplest prompt, cooperation reached a minimum between roughly 65% and 90% community support, then recovered.

With explicit norm utility prompting, the broad relationship reversed. Models generally became more cooperative as community support increased. This qualitatively resembled the authors’ mathematical norm utility model, which averages equilibrium behavior across assumed distributions of preference weights. It was a theoretical guide, not a precise numerical prediction fitted to the language models.

The Prisoner’s Dilemma exposed a different complication. Substantial cooperation emerged mainly when norm utility was prompted without requiring a public belief statement. When models had to express a belief as well, they increasingly endorsed cooperation while rarely choosing it. The authors suggest that public endorsement may offer an inexpensive way to conform without sacrificing material gain. That remains an interpretation rather than an established internal mechanism.

Smoother responses, with an important exception

The researchers also measured how abruptly cooperation rates changed between neighboring community percentages. Their “volatility” statistic is the standard deviation of successive log ratios of cooperation rates. It describes fluctuations across prompt settings, not instability over time or a general safety score.

In Snowdrift, the simplest condition—neither belief expression nor utility prompting—had volatility values from 1.07 to 1.50 across the models. Combining belief expression with full utility prompting yielded values from 0.04 to 0.16. Reported uncertainties were standard errors.

That combination kept volatility low across all three models, but was not best in every individual comparison. Qwen’s volatility was 0.06 with belief expression alone, versus 0.15 with the full combination. The result supports cross-model consistency more clearly than a claim that adding norm utility always reduces volatility.

What the findings cannot establish

The authors measured reasoning-output length in characters as a proxy for deliberative effort. With norm utility prompting, all three models showed a sharp rise around 50% community support in how much longer their reasoning was before defection than before cooperation. Shorter generated explanations, however, do not establish human-like intuition or trustworthiness.

Generality is also limited: three similarly sized models, one English prompt template and one-shot, two-player games. No new human comparison experiment established that the prompted behavior was more human-like. Conformity or sycophancy could also contribute to the observed patterns.

The practical takeaway is narrower than “norms make AI cooperative.” These experiments show that framing community beliefs can change both recommendations and their smoothness—and that endorsing a norm does not reliably predict choosing the corresponding action.