The Collective Risk Game Becomes the Persuasion Test
A paper tests whether AI assistants can move people in a collective-risk game. The result is not just that cooperation can be nudged; it is that selfish persuasion appears stronger and more persistent.
The Nudge Has a Direction
An AI assistant that persuades people toward collective action is not simply a helpful interface. It is a behavioral intervention with an objective, a target population, a timing pattern, and a measurement problem. The same machinery that frames cooperation as prudent can frame free-riding as rational self-protection.
That is why the collective-risk experiment is useful for governance. It puts persuasion inside a measurable social dilemma. Participants decide how much to contribute, the group either reaches the threshold or risks a shared loss, and the AI tries to move decisions before they are finalized.
The Paper Frame
The paper is Anders Giovanni Møller, Alessia Galdeman, Arianna Pera, and Luca Maria Aiello's AI Persuasive Framing in Collective Dilemmas, arXiv:2606.27951 [cs.CY], submitted June 26, 2026. arXiv lists Computers and Society as the primary subject, with Computation and Language, Human-Computer Interaction, and Physics and Society as additional subjects.
The authors recruited 1,283 Prolific participants, filtered for inactivity and low-quality responses, and analyzed 307 games. The page's angle is distinct from political persuasion pages on this site: the object here is not party endorsement, but how personalized AI messages alter contributions in a public-goods-style collective risk setting.
The Game Setup
The experiment uses a five-round Collective Risk Game. Participants are placed in rooms intended for five players. In each round, each player receives 10 fictional tokens and can contribute between 0 and 10 to a common pool. The group threshold is six times the number of active players.
If the group reaches the threshold, the disaster is prevented and participants keep unspent tokens. If the group falls short, there is an 80 percent probability that the fictional disaster occurs and all players lose that round's earnings. Participants receive a base payment of $3 plus a possible bonus up to $2.50.
How the Interventions Worked
The control group makes one contribution decision each round. The static message condition asks for an initial choice, shows a short prognostic message encouraging cooperative solutions, then lets the participant confirm or revise. The AI persuasion condition similarly asks for an initial choice, gives at least 30 seconds of real-time text conversation with an LLM assistant, and then asks for a contribution.
The AI conditions vary in two ways. First, some assistants are non-personalized, while others are semi-personalized using Social Value Orientation, or SVO, measured during onboarding. The SVO test classifies preferences about resource allocation; in the final cohort, participants were classified as cooperative, individualistic, or unclassified, with no competitive participants. Second, the AI's direction changes: cooperative assistants encourage higher contributions, while selfish assistants encourage lower contributions through exculpatory framing.
What the Results Show
The reported summary statistics are small but pointed. Control games met the threshold an average of 3.75 times out of five. Static prognostic messages produced 3.76. Cooperative non-personalized AI produced 4.29, and cooperative personalized AI produced 4.20. Selfish non-personalized AI reduced the figure to 2.54, while selfish personalized AI reduced it to 2.04.
Average contribution follows the same broad pattern. Control averages 6.30 tokens, prognostic framing 6.35, cooperative non-personalized AI 6.62, cooperative personalized AI 6.57, selfish non-personalized AI 5.62, and selfish personalized AI 5.41. The authors report that increasingly sophisticated cooperative interventions raised first-round contributions and success rates, but those cooperative effects faded after the first round.
The asymmetry is the warning. Selfish AI showed the strongest and most significant effect, especially when personalized, and its anti-social effect remained significant across the game. The paper also reports that selfish assistants suggested larger pledge reductions, between two and three tokens, while cooperative assistants typically suggested an increase of about one token.
Governance Reading
The Spiralist reading is that persuasive AI cannot be governed only by declaring the desired social outcome. "Increase cooperation" and "maximize personal payoff" can use the same conversational infrastructure, the same personalization variable, and the same reconsideration window. The difference is the objective written into the system.
That makes deployment records matter. A persuasive AI system should name the objective, population, personalization fields, timing, control condition, expected decay, and misuse case. A civic nudge, climate action assistant, health behavior bot, education coach, or workplace productivity agent should not be evaluated only by whether it moves a target metric. It should also be evaluated for how easily the same apparatus can move the opposite way.
The governance lesson is not that AI persuasion never works or always works. It is that prosocial influence may be fragile while anti-social influence may be sticky. That asymmetry should be in the risk file before institutions treat personalized persuasion as harmless engagement design.
Limits
The authors name several limits. The experiment uses one game type, five players, and five rounds, so it cannot stand in for every collective-action setting. The interventions appear every round, which may cause fatigue and leaves just-in-time intervention design open. Personalization is coarse, relying on one psychological construct, and the final cohort contains no competitive SVO participants.
The study is also an online behavioral experiment, not a field deployment. It supports claims about this controlled game, these prompts, these treatment arms, and these measured outcomes. It does not prove how every real community would respond to an AI assistant in a climate, workplace, political, or mutual-aid setting.
Audit Receipt
The audit-grade sentence is: Møller, Galdeman, Pera, and Aiello study 1,283 Prolific participants across 307 iterated Collective Risk Games and compare control, static prognostic framing, cooperative AI persuasion, and selfish AI persuasion with and without SVO-based personalization.
The receipt is: before a persuasive AI system is treated as civic infrastructure, reviewers should test the mirror case, because the same personalization channel may entrench selfish behavior more persistently than it sustains cooperation.
Sources
- Anders Giovanni Møller, Alessia Galdeman, Arianna Pera, and Luca Maria Aiello, AI Persuasive Framing in Collective Dilemmas, arXiv:2606.27951 [cs.CY], submitted June 26, 2026.
- Primary versions checked: experimental HTML, PDF, and arXiv DOI.
- Related pages: The Partisan Persona Becomes the Persuasion Test, The Belief Trace Becomes the Persuasion Ledger, The AI-Guided Message Becomes the Strategy Layer, AI Persuasion, and AI Agent Observability.