Fostering Collective Intelligence in Human–AI Collaboration: Laying the Groundwork for COHUMAIN

Pranav Gupta,a Thuy Ngoc Nguyen,b Cleotilde Gonzalez,bAnita Williams Woolleyca Gies College of Business, University of Illinois, Urbana-Champaignb Department of Social & Decision Sciences, Carnegie Mellon Universityc Tepper School of Business, Carnegie Mellon University

T

This paper introduces COHUMAIN (Collective Human-Machine Intelligence), a research agenda and framework for studying the complex dynamics of human-AI collaboration. The authors argue that existing research on human-machine interaction is often fragmented across different disciplines, leading to social science models that don't fully account for technology and technical systems that don't consider foreseeable "unexpected" outcomes. COHUMAIN proposes a holistic and interdisciplinary approach to designing and developing sociotechnical systems where humans and AI agents can effectively work together.

​

A central goal of COHUMAIN is to foster collective intelligence (CI), which is a group's ability to solve a wide range of problems. The framework is built upon sociocognitive architectures, which provide the underlying infrastructure for how a complex system, like a human-AI team, perceives, understands, and acts productively.

The paper identifies four core problems that any sociocognitive architecture must address to enable CI:

​

1. Mental States: How individuals perceive their own and others' mental states (e.g., goals, beliefs).

2. Cognitive Resources: How individuals perceive and represent their own and others' specialized knowledge and skills.

​

Two crucial elements for developing these architectures are human-AI trust and Machine Theory of Mind (MToHM). MToHM refers to an AI agent's ability to predict another agent's (human or machine) cognitive state, such as their beliefs, desires, and intentions. This capability is foundational for effective interaction and collaboration.

To further structure their approach, the authors present the Transactive Systems Model of Collective Intelligence (TSM-CI). This model breaks down CI into three interlocking systems:

​

• Transactive Memory System (TMS): Coordinates distributed knowledge and skills, essentially managing "who knows what" within the group. AI can enhance this by helping team members learn new skills or find information quickly.

• Transactive Attention System (TAS): Coordinates the team's focus and priorities to ensure efficient use of collective attention.

• Transactive Reasoning System (TRS): Aligns the team's diverse goals and preferences to facilitate joint decision-making.

​

Finally, the paper advocates for using Instance-Based Learning Theory (IBLT) as a compatible cognitive architecture for developing AI agents within the COHUMAIN framework. IBLT models decision-making based on memories of past experiences and is particularly well-suited for creating AI agents that can interact with humans and develop a Machine Theory of Mind. By integrating these concepts, the authors aim to unlock the full potential of human-machine intelligence