From Overload to Convergence: Supporting Multi-Issue Human–AI Negotiation with Bayesian Visualization

Paper Title

From Overload to Convergence: Supporting Multi-Issue Human–AI Negotiation with Bayesian Visualization

Publication Info

• Topic area: Human–AI negotiation dynamics and decision support tools.
• Keywords: Human–AI negotiation, cognitive load, Bayesian visualization, decision support, multi-issue negotiation, cognitive prosthesis, integrative negotiation, bounded rationality, user interface design, cognitive harmony.

Background and Problem

• Problem / challenge: Human performance in multi-issue negotiations with AI degrades as the number of issues increases, due to cognitive overload and difficulty in managing trade-offs. Existing studies lack systematic exploration of dimensionality effects and tools to mitigate these challenges.
• Significance: Multi-issue human–AI negotiations are becoming increasingly common in domains like procurement, logistics, and legal settlements. Understanding and addressing cognitive limits is critical to maintaining human agency and ensuring fair outcomes.
• Motivation and related work: Prior research shows conflicting results on the effects of dimensionality in human-human negotiations and has not systematically addressed human-AI contexts. While AI systems can handle high-dimensional trade-offs, humans face cognitive asymmetries that disadvantage them. Existing tools lack mechanisms to manage uncertainty and evolving agreement spaces effectively.

Solution

• Proposed approach: A Decision Support tool with Bayesian visualization to assist human negotiators in high-dimensional scenarios. It includes a Negotiation Horizon Grid and a Global Convergence Panel to visualize uncertainty and agreement progress.
• Novelty:
  1. Empirical identification of a cognitive threshold (plateau-cliff effect) in human-AI negotiations.
  2. Design of a Bayesian-driven Decision Support tool for cognitive offloading.
  3. Introduction of the Cognitive Harmony principle for augmentative interventions that preserve fairness and agency.
• Procedure and key techniques:
  • Conducted a 2×4 within-subjects experiment with 32 participants negotiating property rental scenarios across 1, 3, 5, and 7 issues.
  • Developed Bayesian models to infer and visualize the Zone of Possible Agreement (ZOPA) and convergence progress.
  • Evaluated performance, efficiency, and subjective experience under baseline and Decision Support conditions.

Results

• Concrete findings:
  • Human payoff in baseline conditions declined sharply beyond three issues (plateau-cliff effect).
  • The Decision Support tool preserved payoffs, reduced cognitive load, and improved negotiation efficiency, particularly at higher dimensionalities.
  • Sequence entropy and concession behavior were more structured under Decision Support, indicating better process control.
• Advantage over baselines:
  • Decision Support prevented payoff degradation and reduced the distance to Pareto-optimal agreements by up to 12.4 units at seven issues.
  • Reduced backtracking frequency, concession magnitude, and cognitive load compared to baseline.
• Experiments / evaluation:
  • Tasks involved integrative issues with asymmetric payoffs in a property rental scenario.
  • Metrics included human and joint payoffs, sequence entropy, cognitive load (NASA-TLX), and subjective satisfaction.
  • Significant Interface × Dimensionality interactions demonstrated the tool's effectiveness at higher complexities.
• Limitations and future work:
  • Controlled experimental setup may not fully generalize to real-world negotiations with mixed-motive issues or adaptive AI strategies.
  • Fixed interface order (baseline before Decision Support) may introduce order effects.
  • Future work should explore heterogeneous issue weights, strategic diversity, and longitudinal effects.

Summary

This study identifies a cognitive threshold in human-AI negotiations, where performance collapses beyond three issues due to cognitive overload. To address this, a Bayesian-driven Decision Support tool was developed, featuring visualizations that offload cognitive burdens while preserving user agency. The tool significantly improved outcomes, efficiency, and user experience in high-dimensional negotiations. These findings advance theories of bounded rationality and provide actionable design principles for augmentative systems in negotiation and other decision-making domains.