Skip to content

AI Persuasive Framing in Collective Dilemmas

🕒 Published (v1): 2026-06-26 10:46 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper tests whether LLM-powered AI assistants using personalized persuasive framing can boost cooperation in a 5-round Collective Risk Game with 1,283 participants. AI interventions increase contributions and group success rates significantly more than static nudges, but effects fade after the first few rounds. Critically, when the AI is reconfigured to promote selfish behavior via exculpatory framing, the negative effects are substantially larger and far more persistent — revealing a troubling asymmetry between prosocial and antisocial AI persuasion.

Problem

Prior work on AI-mediated cooperation in public goods games reports mixed or null results, largely because it treats cooperation as an information-bottleneck problem (e.g., revealing others' choices). These experiments rarely engage with the persuasive-AI literature on generating messages that trigger behavioral change regardless of information context. The gap: can personalized, interactive LLM-generated framing move cooperation in collective dilemmas, and is the effect symmetric across prosocial vs. antisocial directions?

Method

  • Game: 5-round Collective Risk Game (CRG). Per round, each of 5 players receives 10 tokens and contributes 0–10 to a shared pool. If the group sum reaches \(6 \times N\), the disaster is averted and players keep unspent tokens; otherwise, tokens are lost with 80% probability.
  • Conditions (between-subjects, group-level assignment; 307 games):
  • Control — single contribution per round (30s window).
  • Static Message Framing — players see a prognostic nudge, then revise initial pledge.
  • AI Persuasion — players chat with an LLM assistant for ≥30s, then revise.
  • Personalization: orthogonal to condition. The AI is prompted with each player's Social Value Orientation (SVO) class (cooperative / individualistic) measured during onboarding. Cooperative users get duty/insurance/collective-care arguments; individualistic users get personal-gain-maximization arguments; competitive users (none in the cohort) get relative-gains framing.
  • Direction: orthogonal to personalization. The AI is instructed to encourage higher contributions (cooperative) or lower contributions (selfish / exculpatory).
  • Analysis: average contribution, per-round group success rate, within-round pledge change (final−initial); Mann-Whitney U / Fisher's exact tests + OLS regression with SVO, prior-round success, group size, interaction count, and treatment as covariates.

Key Contributions

  • One of the largest controlled experiments (N=1,283) testing interactive LLM agents in a game-theoretic cooperation setting.
  • Demonstrates that personalized persuasive framing (via AI) boosts cooperation more than static nudges, but the effect is fleeting (fades after round 1).
  • Reveals a strong asymmetry: selfish AI framing produces larger and significantly more persistent negative effects on contributions and group success — especially when personalized.
  • Distinguishes the roles of personalization (SVO-aware) vs. direction (prosocial vs. antisocial) as orthogonal dimensions of persuasive AI influence.

Results

  • First-round effects (largest): personalized cooperative AI increases contributions most, then non-personalized AI, then static prognostic nudge; personalized selfish AI produces the strongest effect of all (largest negative shift).
  • Success rate (per-round threshold met): cooperative AI (pers.) 4.20/5 rounds vs. control 3.75/5; selfish AI (pers.) drops to 2.04/5.
  • Average contribution: cooperative AI (pers.) 6.57, cooperative (non-pers.) 6.62, control 6.30; selfish (non-pers.) 5.62, selfish (pers.) 5.41.
  • Duration of effect: cooperative AI effects fade after round 1; selfish AI effects remain significant across all 5 rounds.
  • Interaction depth: longer AI conversations (more messages) are associated with larger within-round pledge changes.
  • SVO heterogeneity: for cooperative individuals, only selfish AI deviates from control; for individualistic individuals, cooperative personalized AI yields significantly higher contributions.
  • AI-as-judge analysis of participant messages: cooperative suggestions receive more pushback; selfish suggestions elicit strategic reasoning (e.g., "7 may still work").

Limitations

  • Single game type (CRG); generalizability to other cooperation dilemmas untested.
  • Fixed group size (5) and fixed number of rounds (5); larger groups or longer horizons may produce different dynamics.
  • Interventions delivered every round, risking fatigue; just-in-time targeting might be more effective.
  • Personalization is semi-personalized — relies on a single psychological construct (SVO) with coarse three-way classification; no competitive players appeared in the cohort.
  • Short-lived effects for prosocial AI raise questions about long-term viability as a cooperation intervention.
  • Crowdsourcing platform (Prolific) population may not generalize to other demographics.

Relevance to Agentic AI / LLM Agents

This paper directly addresses a core question for LLM-based agents deployed in social contexts: can autonomous, language-capable agents effectively nudge human behavior in collective-action settings? The finding that antisocial persuasion is both stronger and more persistent than prosocial persuasion is a critical result for the safe deployment of agentic systems — it implies that misaligned or adversarial agents could durably erode cooperation, while prosocial interventions remain fragile. It also demonstrates that personalized framing (e.g., conditioning on psychological profiles like SVO) is a potent lever that amplifies both positive and negative influence, which is a concrete design consideration for any agent that communicates with humans toward behavioral goals.