Delegation to artificial agents fosters prosocial behaviors in the collective risk dilemma

Elias Fernández Domingos, Inês Terrucha, Rémi Suchon, Jelena Grujić, Juan C. Burguillo, Francisco C. Santos, Tom LenaertsScientific 2022 · DOI

TL;DR

This paper investigates how delegating decisions to artificial agents affects human cooperation in a high-stakes collective risk dilemma. They found that delegation significantly increases group success by acting as a commitment device, reducing selfish behavior and fear of betrayal, though participants often underestimate AI contributions.

What Problem Does This Paper Address?

As artificial agents (AIs) become increasingly integrated into our daily lives, from home assistants to self-driving cars, understanding their impact on human social interactions is crucial. Specifically, this research addresses the question of whether delegating tasks to AIs in a social context, like a public goods game, promotes or hinders cooperation towards a common good. The authors focus on the "Collective-Risk Dilemma" (CRD), a game where group members must contribute to a shared resource to avoid a significant collective loss, but individual incentives can lead to defection. Previous research on human-to-human delegation has shown mixed results, with some studies indicating it can be used to avoid responsibility or reduce generosity. This paper aims to bridge this gap by experimentally examining the effects of delegating to artificial agents.

Methods

The study employed a series of experimental treatments based on the Collective-Risk Dilemma (CRD). Participants played a game in groups of six for ten rounds, contributing to a common pool with a collective target. Failure to meet the target resulted in a 90% chance of losing their remaining endowment.

The core experimental design involved three treatments compared to a control group playing with only humans:

  • Delegate Treatment: Participants chose one of five pre-defined artificial agents to play on their behalf. These agents exhibited various behaviors, including unconditional contributions (always contributing 0, 2, or 4 units) and conditional behaviors (compensating or reciprocating based on others' actions).
  • Customize Treatment: Participants were given a template and could customize an agent's behavior by setting parameters like a personal threshold and actions based on group contributions. This allowed for a much wider range of agent strategies.
  • Nudge Treatment: This was a hybrid human-agent treatment where groups consisted of three humans and three artificial agents. The agents were selected from those that performed well in the "customize" treatment, with the hypothesis that they might "nudge" human participants towards more cooperative behavior.

Participants' decisions, group outcomes (success/failure), and payoffs were recorded. Post-experiment questionnaires were used to gauge participants' preferences for delegation and their perceptions of agent contributions.

Datasets

Not discussed in the paper. The study relies on data collected from experimental participants.

Key Insights and Ideas

  • Delegation as a Prosocial Catalyst: The most striking finding is that delegating to artificial agents significantly boosts group success in the CRD. This suggests that AIs can act as powerful tools for promoting cooperation, even in high-risk scenarios.
  • Commitment Devices: The paper posits that delegation to AIs functions as a "commitment device." By handing over control, individuals are less tempted by short-term selfish gains and are less susceptible to negative emotions like betrayal aversion, which often hinder cooperation in human-only interactions.
  • The Power of Customization: While simple delegation improved outcomes, allowing participants to customize their agents further enhanced group success. This suggests that having a sense of control or tailoring the AI to one's own preferences increases trust and willingness to delegate.
  • Misperceptions of AI Contributions: A surprising finding is that participants consistently underestimated the contributions of artificial agents, believing they were less cooperative than humans. This perception bias persisted even when agents were actually contributing more.
  • Hybrid Interactions are Tricky: The "nudge" treatment, where humans and agents played together, did not yield the same increase in group success as pure delegation. While some positive effects were observed (e.g., earlier contributions in successful groups), the overall impact was less pronounced, suggesting that hybrid dynamics are more complex.

Key Results

  • Increased Group Success: The "delegate" and "customize" treatments saw a significant increase in the fraction of successful groups, rising from 66.7% in the human-only control to 87% and 87.5%, respectively. The "nudge" treatment showed a similar success rate to the control (68.8%).
  • Payoff Inequality: While group success increased, the variance in individual payoffs also rose across all delegation treatments, indicating a potential increase in payoff inequality.
  • Agent Behavior Bias: In the "delegate" treatment, participants showed a bias towards selecting more cooperative agents (e.g., "always-2," "reciprocal," "compensator") compared to what a baseline model predicted.
  • Participant Preference: Despite the success of delegation, most participants indicated they would prefer to play the game themselves rather than delegate if given the choice again. However, a higher proportion preferred to delegate after the "customize" treatment.
  • Perception vs. Reality: In the "nudge" treatment, 80% of participants in failed groups believed humans contributed more than agents, while agents actually contributed slightly more (+0.33%). In successful groups, this perception shifted, with 48.48% believing humans contributed more, while agents contributed 8.48% higher.

Limitations and Future Work

  • Preference for Agency: A key limitation is that participants, while benefiting from delegation, generally prefer to retain direct control over their actions. This suggests that the full potential of AI for promoting cooperation might be limited by human reluctance to cede agency.
  • Trust and Customization: The authors suggest that customization can increase trust and willingness to delegate, pointing to future work exploring the interplay between delegation, customization, and user satisfaction.
  • Understanding Hybrid Dynamics: The less pronounced effects in the hybrid "nudge" treatment highlight the need for further research into how humans and AIs interact in mixed groups, particularly concerning coordination and social expectations.
  • Perception Bias: The persistent negative perception of AI contributions warrants further investigation into the psychological underpinnings of trust and bias towards artificial agents.
  • Future Research Directions: The authors propose exploring the interplay of social expectations, the embedding of social norms into delegates, and the development of reciprocal rather than purely compensatory strategies in future work. They also note the need to investigate how to balance increased group success with more equitable payoff distributions.

Summary updated

Abstract

Home assistant chat-bots, self-driving cars, drones or automated negotiation systems are some of the several examples of autonomous (artificial) agents that have pervaded our society. These agents enable the automation of multiple tasks, saving time and (human) effort. However, their presence in social settings raises the need for a better understanding of their effect on social interactions and how they may be used to enhance cooperation towards the public good, instead of hindering it. To this end, we present an experimental study of human delegation to autonomous agents and hybrid human-agent interactions centered on a non-linear public goods dilemma with uncertain returns in which participants face a collective risk. Our aim is to understand experimentally whether the presence of autonomous agents has a positive or negative impact on social behaviour, equality and cooperation in such a dilemma. Our results show that cooperation and group success increases when participants delegate their actions to an artificial agent that plays on their behalf. Yet, this positive effect is less pronounced when humans interact in hybrid human-agent groups, where we mostly observe that humans in successful hybrid groups make higher contributions earlier in the game. Also, we show that participants wrongly believe that artificial agents will contribute less to the collective effort. In general, our results suggest that delegation to autonomous agents has the potential to work as commitment devices, which prevent both the temptation to deviate to an alternate (less collectively good) course of action, as well as limiting responses based on betrayal aversion.
BibTeX
@inproceedings{domingos2022delegation,
  author = {Elias Fernández Domingos and Inês Terrucha and Rémi Suchon and Jelena Grujić and Juan C. Burguillo and Francisco C. Santos and Tom Lenaerts},
  title = {{Delegation to artificial agents fosters prosocial behaviors in the collective risk dilemma}},
  booktitle = {Scientific},
  year = {2022},
  doi = {10.1038/s41598-022-11518-9},
}
Figure