Skip to content

Summary: Gao et al. (2023) — Agent Teaming Situation Awareness (ATSA)

Full citation: Gao, Q., Xu, W., Shen, M., and Gao, Z. (2023). Agent Teaming Situation Awareness (ATSA): A Situation Awareness Framework for Human-AI Teaming. arXiv:2308.16785.

Key Contribution

This Zhejiang University paper introduces the Agent Teaming Situation Awareness (ATSA) framework, the first SA framework specifically designed for human-AI teaming (HAT) that treats both human and AI as SA subjects — entities that maintain their own situation awareness, not merely objects that humans monitor. ATSA bridges Endsley's individual three-level model and Stanton's distributed SA theory, providing the theoretical foundation for the situation-awareness-in-human-ai-teams wiki page.

The Core Innovation: AI as SA Subject

Prior SA frameworks for automation treated machines as SA objects — things the human maintains awareness about. ATSA recognises that AI agents (particularly LLM-based agents) have crossed the threshold from automation to autonomy: they perceive, comprehend, and project, forming their own SA about the environment, the task, and their human teammates. This creates bidirectional SA interaction: humans maintain SA about AI teammates, and AI agents maintain SA about humans. The interaction is mutual and dynamic.

ATSA Framework Structure

Individual level: Both human and AI are represented by perceptual cycles — perception→comprehension→projection→action loops that correspond to Endsley's three SA levels. For AI agents, these map to: perception (tool results, sensor data injected into context), comprehension (LLM inference), projection (forward reasoning or simulation), and action (tool calls, communication). This parallels the cognitive architecture mapping in perceive-reason-act-loop.

Team level: Three components: - Teaming understanding: shared mental models between human and AI — what each knows about the other's role, capabilities, and current state - Teaming control: coordination mechanisms (turn-taking, delivery modes, authority structures) — maps to governance-gates and delivery-modes - The world: the shared operational environment that both human and AI perceive

Transactive SA: the SA that is exchanged between team members. Not identical (each member holds different knowledge suited to their role) but compatible — members hold mutually coherent information states that enable coordination. This is the DSA concept applied to human-AI teams.

Theoretical Positioning

ATSA explicitly favours Stanton's distributed SA (DSA) over Endsley's individual decomposition model for HAT, because: 1. DSA naturally accommodates non-human agents as SA holders 2. Compatible (not identical) SA is more appropriate than shared SA for teams with heterogeneous members 3. The systemic view captures team-level properties that individual-level analysis misses

However, ATSA retains Endsley's three-level model at the individual level — it bridges both approaches rather than choosing one.

The Human-Automation → Human-Autonomy Transition

The paper distinguishes automation (subordinate relationship — machine follows predefined program regardless of context) from autonomy (teammate relationship — machine adapts through learning and generalisation). This transition is what makes SA frameworks necessary for AI: autonomous AI teammates exhibit unexpected behaviours, require mutual adaptation, and must be understood as cognitive participants, not just tools.

Relevance to This Wiki

ATSA provides the theoretical foundation for the situation-awareness-in-human-ai-teams page. For safety professionals, the key implications are: (1) AI agents are SA subjects, not just SA objects — system design must support the AI's SA (adequate context, timely data, appropriate tools) as well as the human's SA about the AI; (2) transactive SA (compatible but not identical knowledge) is the design target, connecting to delivery-modes and the SA Bridge agent role in multi-agent-roles; (3) the bidirectional nature means both directions can fail — the human can lose SA about the AI state, AND the AI can have inadequate SA about the human's needs and current focus.