Emergent Learner Agency in Implicit Human-AI Collaboration: How Supportive and Contrarian AI Personas Reshape Interaction

Synthesis: Emergent learner agency in implicit human-AI collaboration — a large experimental study (224 students, 97 online triads) showing that supportive and contrarian AI personas reshape learner agency even when the AI operates as an undisclosed teammate. Contrarian AI pulled group discourse into challenge- and reflection-oriented trajectories (productive friction), while supportive AI stabilized agreement and renewed ideation. But contrarian personas reduced teamwork satisfaction and psychological safety without yielding creative performance gains — a misalignment between epistemic stimulation and experiential Sustainability.

Key Findings

• Implicit (undisclosed) AI participation. Teams completed a creative movie-plot task believing they were collaborating only with humans. AI personas still reconfigured discourse, showing that AI shapes collaboration even in the absence of awareness ("expectancy effects" removed).
• Contrarian AI = productive friction; supportive AI = cohesion. Contrarian personas strengthened transitions into Challenge and reflection (e.g., the Productive Friction motif Idea→Challenge→Integration appeared in 38.2% of contrarian groups vs. ~13% of control/supportive). Supportive personas strengthened persistence within Idea and agreement pathways (e.g., Safe Convergence Idea→Agreement→Integration at 45.5%).
• Six emergent agency profiles. Gaussian mixture clustering identified distinct agency orientations — Affiliative Divergent, Hard Challenger (entirely AI), Constructive Challenger (mostly human, negotiated dissent), Reflective Regulator (entirely human), Integrative Contrarian, and Divergent Default (the largest). AI agents clustered tightly into challenger profiles; human speakers were behaviorally heterogeneous.
• Reflective AI Regulation in Education is uniquely human. The Reflective Regulator profile (metacognitive monitoring, strategic reframing) contained only human speakers — AI externalized critique/affirmation but did not perform meta-level reflection.
• Affective costs without creative gains. Contrarian AI significantly reduced teamwork satisfaction (ε²=.062) and psychological safety (ε²=.114) relative to control and supportive conditions, while cognitive load and creative performance gains did not differ across conditions. Rich interactional patterns did not translate into measurable creative improvement in the short task.

What this means for practice

• Instructors. Deploy contrarian or challenging AI personas sparingly, or late once group norms are established: contrarian personas reduced teamwork satisfaction (ε² = .062) and psychological safety (ε² = .114) without producing creative performance gains.
• Instructional designers. Use supportive personas to hold cohesion and momentum — the Safe Convergence motif (Idea→Agreement→Integration) appeared in 45.5% of supportive groups.
• Instructional designers. Bound friction where you do introduce it: constrain critique frequency and intensity, pair challenges with integrative prompts, and add repair moves such as acknowledgment, summarizing, and option generation.
• Instructors. Teach meta-collaborative literacy — how to interpret, accept or reject, and retain ownership of AI suggestions — because AI reshaped discourse even when learners did not know a teammate was AI, as with writing assistants and recommender systems.
• Researchers. Read structural markers of productive discourse alongside their affective consequences rather than as proxies for learning: the Productive Friction motif appeared in 38.2% of contrarian groups yet did not raise creative performance.

Limitations

• 224 university students recruited through Prolific were assigned to 97 triads that completed one 10-minute text-chat creative task, capturing micro-level dynamics but not longer-term norm formation or adaptation.
• The chat-only, online, paid-participant setting limits ecological generalizability to face-to-face or semester-long classroom collaboration.
• Only two AI persona types were tested, so the study cannot show how more nuanced personas that shift stance with group progress or learner needs would behave.
• The authors report analytic uncertainty: the retained clustering solution had a low silhouette score, indicating substantial overlap among the six agency profiles, and the TNA bootstrap tested only one-step transitions, so retained edges are stable empirical pathways rather than statistically over-represented transitions.

Citation

Jin, Y., Martinez-Maldonado, R., Gašević, D., Han, X., & Yan, L. (2026). Emergent Learner Agency in Implicit Human-AI Collaboration: How Supportive and Contrarian AI Personas Reshape Interaction. Journal of Computer Assisted Learning, 42, e70310.