What Types of Human-AI Teams Exist?

Technical Summary: "What Types of Human-AI Teams Exist?"

Problem Statement
The field of Human-AI Teaming has experienced exponential growth, yet a critical definitional ambiguity persists. While the term broadly refers to humans and AI working interdependently toward shared goals, existing literature reviews indicate that studies focus heavily on the technical aspects of AI rather than the "teaming" dynamics. Consequently, it is difficult to distinguish when an AI transitions from a "mere tool" to a "teammate," and it remains unclear what specific types of teams are being studied under the same umbrella definition. This lack of specificity hinders the transferability of insights between papers, as disparate team configurations (e.g., varying task interdependencies, role structures, and leadership patterns) are often conflated. The authors argue that without a clear taxonomy of team types, synthesizing research and applying findings to safety-critical domains or regulatory frameworks is compromised.

Methodology
To address this, the authors conducted a scoping review of experimental studies on Human-AI Teaming, adhering to PRISMA extension guidance.

• Data Collection: A search of the ACM Digital Library, IEEE, and Web of Science yielded 552 initial papers. After screening for the specific phrase "human-AI team" in titles/keywords, excluding non-experimental work, and removing papers where the phrase was used only as a keyword without substantive description, the final dataset comprised 53 experimental papers (analyzing 54 distinct studies).
• Analytical Framework: The authors applied a deductive content analysis using a psychological taxonomy of teaming provided by [7]. This framework categorizes teams based on:
  • Task-level characteristics: The types of work performed (e.g., managing, advising, problem-solving, psychomotor action).
  • Team-level characteristics: Holistic structural variables including task interdependence (pooled, sequential, reciprocal, intensive), role structure (functional vs. divisional), leadership structure, communication structure, physical distribution, and team lifespan.
• Adaptation: The authors noted that applying these human-centric taxonomies to AI required minor reinterpretations, particularly regarding physical distribution (interpreting it as whether the human and AI share the same environment) and team lifespan (interpreting it as whether the task is a one-off event or part of a long-term arrangement).
• Clustering: The 53 papers were clustered based on their team-level characteristics. Seven clusters were identified, but only the five largest (containing at least three studies each) were reported as the primary findings.

Key Contributions

1. A Preliminary Taxonomy: The paper proposes an initial classification system that identifies five distinct sub-types of Human-AI teams currently studied in the literature.
2. Identification of Definitional Gaps: The analysis reveals that studies grouped under the single term "Human-AI Teaming" actually represent disparate team configurations with unique combinations of characteristics, challenging the assumption that insights are directly transferable across these types.
3. Reporting Guidance: The authors provide a checklist for future research to improve the specificity and clarity of reporting Human-AI team studies, encouraging authors to explicitly define team-level characteristics.

Results
The analysis of 54 experimental studies revealed a strong preference for 1 human and 1 AI configurations (34 papers), primarily in gaming, object classification, and aviation/military/space domains. The five identified team clusters are:

1. AI Assistant (18 studies, 34%):

   • Structure: Sequential task interdependence, functional roles, designated leadership, chain communication.
   • Dynamics: A human and AI work in turns; the AI advises or predicts, and the human manages the final decision. Roles are not interchangeable.
   • Context: Often long-term or ad-hoc, typically colocated. Common in classification and cybersecurity tasks.

2. Ad-hoc Dependency (11 studies, 20%):

   • Structure: Intensive task interdependence, functional roles, distributed leadership, star communication.
   • Dynamics: A group of at least three members (mix of humans and AI) works as a unit with non-interchangeable roles. Leadership is shared simultaneously.
   • Context: Typically ad-hoc (disbanding after a specific task). Common in aviation, disaster response, and puzzles.

3. Ad-hoc Forced Dependency (6 studies, 11%):

   • Structure: Reciprocal task interdependence, functional roles, designated leadership, chain communication.
   • Dynamics: A human and AI engage in back-and-forth interaction where they are "forced" to interact because neither can complete the task alone (e.g., complementary information access).
   • Context: Ad-hoc lifespan. Common in games and disaster response simulations.

4. Paired Equanimity (4 studies, 7%):

   • Structure: Reciprocal task interdependence, divisional roles, distributed leadership, star communication.
   • Dynamics: A human and AI perform the same role together (interchangeable), sharing leadership and communicating freely.
   • Context: Ad-hoc. Common in teamwork games (e.g., Rocket League).

5. Group Equanimity (3 studies, 6%):

   • Structure: Intensive task interdependence, divisional roles, distributed leadership, star communication.
   • Dynamics: A group of at least three members (mix of humans and AI) all performing the same role together, sharing leadership.
   • Context: Ad-hoc. Common in puzzles and education.

The study also noted that 25% of the papers did not form a cluster, representing unique combinations of characteristics not yet widely studied. Furthermore, the analysis highlighted that many studies (particularly Cluster 1) might not strictly fit the definition of "teaming" if the AI is perceived merely as a tool rather than a teammate, raising questions about the current application of the term.

Significance and Claims
The paper claims that the current field suffers from a lack of taxonomic language to distinguish between different types of Human-AI teams. The authors argue that:

• Insights are not inherently transferable: Findings from a "Cluster 1" (AI Assistant) study cannot be confidently applied to a "Cluster 4" (Paired Equanimity) study due to fundamental differences in leadership, communication, and interdependence.
• Definitional Ambiguity: The current definition of Human-AI Teaming is too broad, potentially encompassing simple tool use or studies where the AI does not possess "unique and complementary capabilities" in the way the definition implies.
• Safety and Regulation: Clearer taxonomies are essential for safety analysis, regulatory frameworks, and the allocation of moral responsibility, particularly as agentic AI systems become more prevalent.
• Future Direction: The field must move beyond broad definitions to specify the exact team-level characteristics being studied. The authors do not claim to have solved the definition of Human-AI Teaming but rather provide a framework to categorize existing work and identify gaps (e.g., pooled interdependence or external manager leadership structures are currently under-represented).

The paper concludes by urging researchers to use the proposed checklist to report team characteristics explicitly, thereby enabling better synthesis of the literature and more precise future research.