Document Type
Poster
Publication Date
Fall 2025
Abstract
Today, large language models are often applied as smart agents who are capable of communicating, negotiating and making decisions together in the same space. In this work, I designed a test scenario where three different AI agents, having different goals, a cooperative helper, a competitive maximizer, and a deceptive agent, are sharing resources in a number of rounds. I would like to observe how the different goals of the agents affect the emotions of the agents. Each time the agents meet, I interpret their communications to assign them numerical values that measure the intensity of their perceived positivity, excitement, or control. I do that based on the Valence-Arousal-Dominance technique. In addition, I use the data to determine their emotional patterns. For instance, I measure the stability of their emotional expressions, analyze shock bursts in their negativity or excitement, and note when one agent manages to convince and trick the other. To demonstrate the influence flow between the two, I have used a diagram to illustrate the impact they have on each other. By integrating the measurement of emotions and the study of networks, this test relates the diverse emotional trajectories that emerge from the goal systems of each agent. The significance of this work is found at the intersection of important notions such as modeling, networks, persuasion, and programming human-like behaviors and lies within the development of trustworthy AI. Emotional tracking can be used as an early detector of malfunctioning and persuader agents.
Recommended Citation
Idowu, Godwin, "Emotional Stability and Deception in Multi-Agent AI Negotiations Tracking Emotional Drift, Persuasion and Influence Using Valence-Arousal-Dominance" (2025). IPHS 200: Programming Humanity. Paper 91.
https://digital.kenyon.edu/dh_iphs_prog/91
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.