Humans program artificial delegates to accurately solve collective-risk dilemmas but lack precision

Inês Terrucha, Elias Fernández Domingos, Rémi Suchon, Francisco C. Santos, Pieter Simoens, Tom LenaertsProceedings of the National Academy of Sciences 2025 · DOI

TL;DR

This study investigates how humans delegate strategic decisions to artificial agents in collective-risk dilemmas. While delegation encourages higher contributions to public goods, it introduces precision errors that hinder group success. The findings suggest that artificial delegates can help sustain cooperation in the long term, but achieving optimal outcomes requires humans to refine their agents' programming.

What Problem Does This Paper Address?

As autonomous machines become more integrated into our lives, strategic decisions with collective consequences are increasingly being delegated to artificial agents. This raises a critical question: what gets lost in translation when humans delegate to algorithms, particularly in situations requiring cooperation for collective benefit? The paper focuses on the "collective-risk dilemma" (CRD), a game where individuals must contribute to a public good that has a target threshold. If the threshold is met, everyone benefits; if not, everyone risks losing their endowment. The study explores how delegation affects contributions, group success, and individual behavior in this context, especially when participants can reprogram their agents after experiencing initial outcomes.

Methods

The researchers conducted a series of online behavioral experiments using a modified Collective Risk Dilemma (CRD). Participants were divided into groups of four and played the game over 10 rounds. The experiment employed a 2x2 factorial design, varying two key factors:

  1. Delegation vs. No-Delegation: Participants either made their contributions directly or programmed an "artificial agent" to make decisions on their behalf.
  2. Action Space Granularity: Participants could choose from either 3 discrete contribution levels (0, 2, 4) or 5 discrete levels (0, 1, 2, 3, 4) per round.

A crucial element was the "surprise restart" of the game. After playing Game 1, participants were invited to play Game 2 with the same group, with the opportunity to reprogram their agents if they had delegated. This design aimed to assess how prior experience (success or failure) and the ability to adjust strategies influenced subsequent behavior. Behavioral profiles were identified using K-Means clustering on the round-by-round action sequences.

Datasets

The study utilized data collected from online behavioral experiments. The participants were recruited through Prolific. The anonymized .csv data and supporting code have been deposited in Zenodo (DOI: https://doi.org/10.5281/zenodo.11033297). No pre-existing datasets were used for training or evaluation.

Key Insights and Ideas

  • Delegation as a Commitment Device: Delegation, even with the option to reprogram, encourages higher contributions to the public good, especially in constrained action spaces. This suggests delegates act as a form of commitment, pushing individuals to contribute more even if they've experienced past failures or inequality.
  • Precision vs. Accuracy Trade-off: While delegation increases the accuracy of aiming for the collective target (i.e., contributing more consistently), it introduces precision errors in the agent's algorithm. This means groups get closer to the target but often miss it due to the agent's programming, leading to lower success rates compared to non-delegation groups in certain conditions.
  • Resilience to Failure: Individuals who delegate appear more resilient to past failures. Unlike non-delegation groups, where failure can lead to reduced contributions in subsequent rounds, delegating groups tend to increase their contributions after experiencing setbacks, suggesting a focus on correcting mistakes rather than punishing group members.
  • Granularity Matters: The number of available choices (action space granularity) significantly impacts behavior. With more choices, individuals in non-delegation settings can be more nuanced and avoid outright zero contributions, potentially sustaining cooperation. However, a constrained action space can amplify signals of selfish behavior, leading to a breakdown in cooperation.
  • Programmed Behavior vs. Observed Behavior: Interestingly, while participants programmed their agents with conditional strategies (reacting to others' actions), the agents' actual behavior often appeared fixed due to the limited parameters and the nature of the CRD. This disconnect between the intention behind programming and the emergent behavior is a key finding.

Key Results

  • Increased Contributions: In the constrained 3-action space, delegation groups contributed significantly more to the Public Account in both Game 1 and Game 2 compared to no-delegation groups.
  • Lower Success Rates with Delegation (in some cases): For the 5-action space, groups with delegation were significantly less successful than those without delegation in both games, despite comparable contributions. This highlights the precision issue.
  • Higher In-Group Inequality: Delegation treatments consistently showed higher variance in private accounts within groups, indicating greater inequality in individual contributions.
  • Behavioral Adaptation: Individuals in delegation treatments were significantly more likely to change their behavior between Game 1 and Game 2, particularly increasing contributions after group failure.
  • Behavioral Profiles: Four distinct behavioral profiles were identified: early, high, fair, and low contributors. Delegation treatments had a larger proportion of "high" contributors and fewer "early" contributors compared to no-delegation.

Limitations and Future Work

  • Precision Errors: The primary limitation identified is the lack of precision in artificial delegates' algorithms, which prevents groups from achieving optimal success even when contributions are high. The authors suggest that actual success requires humans to learn to adjust their agents' algorithms more effectively.
  • Granularity Effects: The study notes that while a larger action space can facilitate cooperation in non-delegation settings, the interaction between delegation and granularity needs further exploration.
  • Future Research: The authors suggest further research into additional mechanisms to improve the precision of artificial delegates. They also propose exploring how the transparency of agent programming might influence human behavior and trust, potentially mitigating feelings of betrayal and fostering more effective cooperation. The "cold" versus "hot" decision-making states of participants when programming agents also warrants further investigation.

Summary updated

Abstract

In an era increasingly influenced by autonomous machines, it is only a matter of time before strategic individual decisions that impact collective goods will also be made virtually through the use of artificial delegates. Through a series of behavioral experiments that combine delegation to autonomous agents and different choice architectures, we pinpoint what may get lost in translation when humans delegate to algorithms. We focus on the collective-risk dilemma, a game where participants must decide whether or not to contribute to a public good, where the latter must reach a target in order for them to keep their personal endowments. To test the effect of delegation beyond its functionality as a commitment device, participants are asked to play the game a second time, with the same group, where they are given the chance to reprogram their agents. As our main result we find that, when the action space is constrained, people who delegate contribute more to the public good, even if they have experienced more failure and inequality than people who do not delegate. However, they are not more successful. Failing to reach the target, after getting close to it, can be attributed to precision errors in the agent's algorithm that cannot be corrected amid the game. Thus, with the digitization and subsequent limitation of our interactions, artificial delegates appear to be a solution to help preserving public goods over many iterations of risky situations. But actual success can only be achieved if humans learn to adjust their agents' algorithms.
BibTeX
@inproceedings{terrucha2025humans,
  author = {Inês Terrucha and Elias Fernández Domingos and Rémi Suchon and Francisco C. Santos and Pieter Simoens and Tom Lenaerts},
  title = {{Humans program artificial delegates to accurately solve collective-risk dilemmas but lack precision}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2025},
  doi = {10.1073/pnas.2319942121},
}

Loading related content...

Figure