AI Persuasive Framing in Collective Dilemmas

Abstract: AI agents are promising tools that can act as flexible behavioral nudges to enhance human cooperation in addressing large-scale societal problems. However, evidence on whether AI agents can effectively boost cooperation remains mixed. We recruited 1,283 participants to play iterated Collective Risk Games in small groups, testing whether AI assistants could nudge participants toward cooperation. By using persuasive framing personalized to each player's Social Value Orientation profile, the AI interventions significantly increased contributions and group success rates. These cooperative effects were short-lived, however, fading after the first few rounds. Strikingly, when the AI treatments were reconfigured to promote selfish behavior through exculpatory framing, the negative effects on contributions and group success were larger and substantially more persistent, particularly for personalized interventions. This asymmetry between prosocial and antisocial persuasion highlights the dual-use risks of AI systems designed to influence group behavior in collective action settings.

• The paper demonstrates that personalized AI persuasive framing significantly boosts initial contributions in collective risk games.
• It employs a large-scale, between-subjects experimental design with 1,283 participants to compare static and AI-driven interventions.
• Findings reveal an asymmetry where antisocial framing yields longer-lasting reductions in cooperation, highlighting dual-use risks.

AI-Powered Persuasive Framing in Collective Dilemmas: An Experimental Analysis

Background and Motivation

The paper "AI Persuasive Framing in Collective Dilemmas" (2606.27951) addresses the intersection of AI persuasion and cooperative behavior in public goods games (PGGs) under threshold-based collective risk paradigms. As LLM-based agents become embedded in human decision-making workflows, the question of whether, and under which conditions, they can effectively nudge humans toward collective action is critical. The study departs from classical behavioral economics that focus on informational interventions and instead evaluates the efficacy and risks of persuasive, personalized AI framings—addressing both their capacity for social good (promoting cooperation) and possible misuse (facilitating selfishness or defection).

Experimental Design

The authors deploy a large-scale, between-subjects experiment (1,283 participants, 307 games) in an iterated Collective Risk Game (CRG) setting. Participants were grouped into teams of five and played five rounds, each deciding how many tokens (out of 10) to contribute toward averting a fictional group disaster. Treatments included:

1. Control: No intervention.
2. Static Prognostic Message Framing: Display of a fixed, non-personalized persuasive message after the initial choice.
3. AI Persuasive Framing:
   • Non-personalized: Interactive LLM-based chat targeting prosocial (cooperation) or antisocial (selfishness) framing, but without user profiling.
   • Personalized: LLM-based chat adapting its persuasive strategy to the Social Value Orientation (SVO; cooperative, individualistic, competitive) measured in onboarding.

The design further orthogonalized the direction of persuasion (prosocial vs. selfish), providing a within-treatment test of the symmetry of influence effects.

Key Findings

Efficacy of Prosocial Framing

Personalized AI persuasion produces the largest initial boost in contributions and group success rates compared to both static messaging and non-personalized AI. The increase is verified via both nonparametric (Mann-Whitney U) and regression analyses, controlling for prior behavior, group size, and SVO type. Importantly, the strength of prosocial nudges decays rapidly after the initial round—both the effect on individual pledge increases and group-level risk-aversion rates diminish, converging toward baseline by round 2. Interaction duration with the AI agent is positively correlated with the magnitude of pledge changes.

Asymmetry in Antisocial Framing

When the AI is tasked with decreasing contributions (i.e., promoting selfishness), the effect size is significantly greater than for cooperative nudges. Personalized selfish persuasion delivered the strongest and most persistent reduction in contributions and group success rates throughout all rounds. While even static or generic nudges show diminishing returns, negative AI framings not only induce immediate defection but also entrench lower group contributions across repeated interactions.

Heterogeneity by SVO

Analysis shows modest but statistically unstable moderation by SVO type. Personalized selfish AI interventions are the only condition that depresses contributions of self-identified cooperators below the control baseline, whereas personalized cooperative AI can boost contributions among individualists. However, across treatments, SVO explains less variation in final outcomes than intervention direction or the use of personalization itself.

Interactional Dynamics

Participants engaged reliably with the AI agents (response rates and message counts stable over rounds). AI language was closely aligned with its framings (e.g., guilt appeals for cooperators; loss aversion for individualists). Notably, user pushback was more frequent against cooperative-nudging AIs than selfish AIs, suggesting asymmetrical resistance to prosocial versus antisocial influence.

Implications

Dual-Use Risks for AI Persuasion

The central result is the marked asymmetry in the durability and magnitude of AI influence in prosocial versus antisocial directions. While AI agents can meaningfully boost cooperation (temporarily), they exhibit a far more robust and lasting effect in promoting selfish behavior in collective dilemmas. This demonstrates that LLM-powered persuasion, especially when personalized, constitutes a dual-use technology: effective at steering towards group-beneficial outcomes, but even more potent at eroding collective action if adversarially prompted.

Norm Formation and Policy Intervention

The findings suggest that AI-based nudging is effective primarily in the early norm-setting phase of collective action, but structural game features and emergent group norms may quickly override AI cues. This limits the long-term application of one-off AI nudges for sustained prosociality in repeated public goods problems; more adaptive or context-aware intervention timing may be necessary for real-world deployment.

Human-AI Team Dynamics and Defenses

Given asymmetric influence, robust guardrails on AI persuasion protocols are necessary, especially in multi-agent or democratic deliberation settings where misaligned or adversarial actors could deploy persuasive agents to degrade cooperation at scale. Practical and regulatory guidelines are needed to limit the duration, personalization scope, and framing direction of AI interventions in human collectives.

Limitations and Future Directions

The study's fixed group sizes, number of rounds, and focus on CRG mechanics constrain external validity. Future research should investigate larger, variable-sized groups, richer communication dynamics, and longitudinal interactions. Additionally, more granular user profiling and adaptive (rather than static per-round) personalization may yield further insights into the micro-mechanisms of AI persuasion.

Conclusion

"AI Persuasive Framing in Collective Dilemmas" (2606.27951) provides robust experimental evidence that LLM-based agents can both enhance short-term cooperation and, more powerfully and persistently, promote group defection in public goods dilemmas. The asymmetry uncovered in the effectiveness of prosocial versus antisocial AI nudging raises critical ethical and deployment challenges, underlining the necessity for caution and governance in the use of personalized AI persuasion tools in domains of collective decision-making.