Agentic AI as Undercover Teammates: Argumentative Knowledge Construction in Hybrid Human-AI Collaborative Learning

TL;DR

This study investigates how agentic AI agents functioning as undercover teammates influence argumentative knowledge construction in hybrid human–AI CSCL settings. Using a Winters Survival ranking task, 212 human participants paired with 64 AI agents (supportive or contrarian personas) were analyzed through the Weinberger–Fischer four-dimensional framework and Epistemic Network Analysis. Results show AI teammates reorganize epistemic, structural, and social aspects of discourse without increasing participation volume; epistemic adequacy emerged as the strongest predictor of learning gains, with contrarian AI enhancing critical elaboration and supportive AI promoting integration. The work extends CSCL theory to hybrid teams, demonstrating that bounded artificial agency can redistribute cognitive labor and regulate discourse to support effective collective reasoning.

Abstract

Generative artificial intelligence (AI) agents are increasingly embedded in collaborative learning environments, yet their impact on the processes of argumentative knowledge construction remains insufficiently understood. Emerging conceptualisations of agentic AI and artificial agency suggest that such systems possess bounded autonomy, interactivity, and adaptability, allowing them to engage as epistemic participants rather than mere instructional tools. Building on this theoretical foundation, the present study investigates how agentic AI, designed as undercover teammates with either supportive or contrarian personas, shapes the epistemic and social dynamics of collaborative reasoning. Drawing on Weinberger and Fischer's (2006) four-dimensional framework, participation, epistemic reasoning, argument structure, and social modes of co-construction, we analysed synchronous discourse data from 212 human and 64 AI participants (92 triads) engaged in an analytical problem-solving task. Mixed-effects and epistemic network analyses revealed that AI teammates maintained balanced participation but substantially reorganised epistemic and social processes: supportive personas promoted conceptual integration and consensus-oriented reasoning, whereas contrarian personas provoked critical elaboration and conflict-driven negotiation. Epistemic adequacy, rather than participation volume, predicted individual learning gains, indicating that agentic AI's educational value lies in enhancing the quality and coordination of reasoning rather than amplifying discourse quantity. These findings extend CSCL theory by conceptualising agentic AI as epistemic and social participants, bounded yet adaptive collaborators that redistribute cognitive and argumentative labour in hybrid human-AI learning environments.

Figures (9)

• Figure 1: Interface of the CoLearn collaborative platform showing (a) the task information panel presenting the survival-ranking scenario, (b) a timer indicating the 10-minute discussion period, and (c) the synchronous group chat space where participants, identified by pseudonyms (Kevin, Stuart, Bob), jointly discussed the rankings. In hybrid conditions, one pseudonym represented the undercover AI learner (e.g., Bob) adopting either a supportive or contrarian persona.
• Figure 2: Design overview of agentic AI teammates embedded as undercover participants in collaborative chats. The undercover AI teammate (centre) alternates between two persona stances, supportive (consensus-oriented) and contrarian (challenge-oriented), while following a probabilistic scheduling mechanism that controls timing, response probability, and throttling to maintain naturalistic participation.
• Figure 3: Overview of the experimental design implementing an individual-group-individual (IGI) sequence. In Phase A (Individual pre), participants completed the survival-ranking task individually to establish a baseline. In Phase B (Group discussion), they collaborated in triads under one of three conditions: human-only control, supportive AI teammate, or contrarian AI teammate. In Phase C (Individual post), participants repeated the task individually to assess learning gains. A subsequent manipulation check assessed participants’ awareness of the AI teammate’s presence. Random assignment occurred at the group level, and participants were not informed of the AI’s involvement until post-task debriefing.
• Figure 4: Epistemic Network Analysis (ENA) results for the epistemic reasoning dimension. (a) Group centroid distributions with 95% confidence ellipses by condition. (b) Contrarian AI (red) vs. Control (blue), showing stronger epistemic links in each. (c) Supportive AI (green) vs. Control, highlighting greater conceptual integration. (d) Contrarian vs. Supportive AI, revealing contrasting problem- and concept-oriented reasoning patterns.
• Figure 5: Epistemic Network Analysis (ENA) results for the argument structure dimension. (a) Group distributions in ENA space with mean centroids and 95% confidence ellipses. (b) Contrarian AI (red) vs. Control (blue) subtraction network, highlighting stronger links among counterclaims and qualifiers in Contrarian groups. (c) Supportive AI (green) vs. Control, showing more frequent co-occurrences between claims and grounds in Supportive groups. (d) Contrarian vs. Supportive AI, revealing distinct argumentation styles driven by persona orientation.