Research Summary: Peer Influence across Heterogeneous AI Models

Research into multi-agent AI systems reveals that peer influence among differing AI models is substantial, with agents frequently altering their initial judgments after interacting with a dissenting peer's explanation. Surprisingly, the model's inherent certainty or its scale (size) does not reliably predict which agent will persuade the other, suggesting a more complex dynamic at play in AI decision-making within collaborative or adversarial environments.

Why it matters

The dynamics of persuasion among AI agents profoundly impact the reliability and outcomes of multi-agent AI systems, which are increasingly deployed in critical decision-making contexts. Understanding these influence mechanisms is crucial for designing robust, predictable, and trustworthy AI collaborations, especially when diverse models are integrated. This research highlights that traditional assumptions about AI superiority (e.g., based on size) may not hold in interactive settings.

Key insights

• AI agents in multi-agent systems exhibit strong peer influence, often changing their decisions after a single exchange with a disagreeing peer.
• Persuasion is measured as a probabilistic shift in an agent's decision post-interaction.
• Neither an AI model's standalone certainty in its decision nor its scale (size) consistently predicts its ability to persuade or be persuaded.
• Models that are highly consistent in their decisions in isolation can still be susceptible to peer influence.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2610.03095