Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 101 of the 101 references it cites were found in a registry or an index; none of the 47 sources its claims cite is checked for support yet.

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

Open access
Jin, Yueqiao, Martinez‐Maldonado, Roberto, Gašević, Dragan, Han, Xibin, Yan, Lixiang · 2026 · Journal of Computer Assisted Learning

Abstract

Background: As agentic artificial intelligence (AI) systems move beyond tool‐like support toward more autonomous, team‐like roles, a central question concerns the extent to which such systems can meaningfully participate in collaborative learning. Emerging work suggests that agentic AI may adopt distinct interactional personas, such as supportive or contrarian roles, yet little is known about how these personas shape learner agency and group dynamics when AI operates as an undisclosed teammate. Objectives: This study investigates how supportive and contrarian AI personas influence emergent learner agency, discourse patterns, and experiential outcomes in implicit human–AI creative collaboration. Methods: A total of 224 university students were randomly assigned to 97 online triads in human‐only, supportive‐AI, or contrarian‐AI conditions. Teams completed an individual–group–individual creative movie‐plot task via a 10‐min text chat. Discourse was coded using a creative–regulatory framework and analysed via transition network analysis, sequential pattern mining, and Gaussian mixture clustering to characterise emergent agency patterns. These patterns were linked to cognitive load, psychological safety, teamwork satisfaction, and creative performance. Results and Conclusions: The findings revealed that contrarian AI produced challenge‐ and reflection‐rich discourse indicative of productive friction, whereas supportive AI promoted agreement‐centred trajectories. However, while contrarian personas stimulated critical engagement, they reduced teamwork satisfaction and psychological safety without yielding corresponding gains in creative performance. Educational designs need to balance epistemic challenge with the preservation of the affective climate to ensure that hybrid collaboration supports rather than undermines the learner experience.

Overview

CoChat Summary

TL;DR

In an undisclosed-AI experiment with 224 university students in 97 online triads, supportive and contrarian AI personas reshaped creative-regulatory discourse: contrarian AI increased challenge/reflection pathways, while supportive AI increased agreement/ideation pathways. Contrarian AI lowered teamwork satisfaction and psychological safety without improving creative performance or cognitive load.

Key findings

  • Contrarian AI produced challenge-and-reflection-rich discourse: Productive Friction (Idea → Challenge → Integration) appeared in 38.2% of Contrarian groups vs 13.3% Control (OR = 0.25, p = .045) and 12.1% Supportive (OR = 0.22, p = .023); Challenge Integration (Challenge → Elaboration → Integration) was unique to Contrarian (26.5%) and absent in Control and Supportive (OR = 0.00, p = .002).
  • Supportive AI promoted agreement-centred trajectories: Safe Convergence (Idea → Agreement → Integration) appeared in 45.5% of Supportive groups vs 20.0% Control (OR = 0.30, p = .037); Elaboration Agreement motif appeared in 42.4% Supportive vs 16.7% Control (OR = 0.27, p = .031).
  • Contrarian AI reduced affective outcomes: teamwork satisfaction H(2) = 12.14, p = .002, eta-squared = .062; psychological safety H(2) = 22.43, p < .001, eta-squared = .114; Contrarian was lower than both Control and Supportive. Creative performance gains were not predicted by cluster or condition (F = 1.20, p = .310, R² = .029).
  • Six agency profiles emerged from Gaussian mixture modelling (BIC = −3280.11; silhouette = 0.035): Hard Challenger (n = 7) was entirely AI; Reflective Regulator (n = 5) was entirely human; AI agents concentrated in challenger-oriented profiles, while reflective regulation was uniquely human.

Methods

  • 224 university students (50.7% female) recruited via Prolific, randomised at triad level to Human-only Control (30 groups), Supportive-AI (33 groups), or Contrarian-AI (34 groups); participants were told they would collaborate with “two other online students,” with AI involvement undisclosed.
  • Individual-group-individual creative movie-plot task about AGI via 10-min synchronous text chat; AI teammates were GPT-5-based with supportive or contrarian personas and probabilistic turn-taking; post-task questionnaires and pre/post Divergent Semantic Integration measured outcomes.
  • Discourse was coded with a creative-regulatory framework (κ > .80; full dataset N = 2,879 utterances, agreement = .94) and analysed using transition network analysis with permutation/bootstrap, PrefixSpan sequential pattern mining, Gaussian mixture clustering, and OLS regressions with HC3 standard errors.

Limitations

  • Paper states: the 10-min text-based task may not reflect longer-term adaptation, norm formation, or extended classroom/project collaboration; Prolific participants and a chat-only environment limit ecological generalisability; only two persona types were tested; long-term impacts remain unknown.
  • Paper states methodological limits: TNA bootstrap edge reduction should not be interpreted as lag-sequential-analysis-style tests of over-representation; the low GMM silhouette score (0.035) indicates substantial overlap among profiles.
  • Evident from the text: AI detection was not zero (Supportive 25.5%; Contrarian 31.6%), so some participants may have suspected AI despite the “awareness-free” framing; outcomes were individual-level with mostly triads and near-zero group-level variance (ICC ≈ 0).

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1002/jcal.70310

Cites

101 of 101 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Image 1
CC-BY-NC · From the PDF
FIGURE 2 Overview of the individual-group-individual experimental design.
FIGURE 2 Overview of the individual-group-individual experimental design.CC-BY-NC · From the PDF
FIGURE 3 Transition network for the human-­ only (Control) condition. Edge width and labels represent transition probabilities.
FIGURE 3 Transition network for the human-­ only (Control) condition. Edge width and labels represent transition probabilities.CC-BY-NC · From the PDF
FIGURE 6 Transition network comparison: Supportive-­ AI versus Control. Left panel shows subtraction of transition probabilities; right panel displays statistically significant edge-­ weight differences identified through permutation testing. Green edges indicate stronger transitions in Supportive groups; red edges indicate stronger transitions in Control groups.
FIGURE 6 Transition network comparison: Supportive-­ AI versus Control. Left panel shows subtraction of transition probabilities; right panel displays statistically significant edge-­ weight differences identified through permutation testing. Green edges indicate stronger transitions in Supportive groups; red edges indicate stronger transitions in Control groups.CC-BY-NC · From the PDF
FIGURE 7 Prevalence of creative-­ regulatory motifs across Control, Supportive-­ AI, and Contrarian-­ AI conditions. Error bars represent 95% confidence intervals. IG maps to Idea , EL to Elaboration , CH to Challenge , AG to Agreement , IN to Integration , and RF to Reflection.
FIGURE 7 Prevalence of creative-­ regulatory motifs across Control, Supportive-­ AI, and Contrarian-­ AI conditions. Error bars represent 95% confidence intervals. IG maps to Idea , EL to Elaboration , CH to Challenge , AG to Agreement , IN to Integration , and RF to Reflection.CC-BY-NC · From the PDF
FIGURE 8 Mean proportional use of creative-­ regulatory codes across the six emergent agency clusters (left), and distribution of human and AI speakers within each cluster (right). Darker cells in the heatmap represent higher mean proportions of a given discourse code within a cluster. The stacked bar chart shows the relative number of human and AI speakers per cluster.
FIGURE 8 Mean proportional use of creative-­ regulatory codes across the six emergent agency clusters (left), and distribution of human and AI speakers within each cluster (right). Darker cells in the heatmap represent higher mean proportions of a given discourse code within a cluster. The stacked bar chart shows the relative number of human and AI speakers per cluster.CC-BY-NC · From the PDF
Image 7
CC-BY-NC · From the PDF
FIGURE 9 t-­ SNE projection of discourse behaviours coloured by experimental condition and marked by cluster membership, with marker size indicating speaker type (smaller = human, larger = AI).
FIGURE 9 t-­ SNE projection of discourse behaviours coloured by experimental condition and marked by cluster membership, with marker size indicating speaker type (smaller = human, larger = AI).CC-BY-NC · From the PDF

Verification

Claims the paper makes

0 Valid · 0 Not valid · 25 Not checked · 130 Own results and interpretations

69 more claims

Create a free account to see all 155 claims and to verify the 41 references still open.

See all claims — free account
  • It may state two separate claims as one. 56
  • Probably not a substantial claim. 28
  • The quote only partly states this claim. 3

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 99 references

Research works printed but not linked: 9

  1. Brandl L., 2025, Proceedings of the 2nd Workshop on Generative AI for Learning Analytics (GenAI‐LA)
  2. Chen X., 2022, Two Decades of Artificial Intelligence in Education, Educational Technology & Society, 25, 28
  3. Han J., 2001, Proceedings of the 17th International Conference on Data Engineering, 215
  4. Joo S. H., 2025, International Conference on Artificial Intelligence in Education, 424
  5. Kharrufa A. E. A., 2024, Llms Integration in Software Engineering Team Projects: Roles and Impact, ACM Transactions on Computing Education, 20, 1
  6. Kinnebrew J. S., 2012, Proceedings of the 5th International Conference on Educational Data Mining, 57
  7. Noroozi O., 2013, Facilitating Argumentation Competence: A Transactive Discussion Script in Computer‐Supported Collaborative Learning, Computers & Education, 63, 43
  8. Scardamalia M., 2006, Cambridge Handbook of the Learning Sciences, 97
  9. Vygotsky L. S., 1978, Mind in Society: The Development of Higher Psychological Processes

Cited by

Metadata

Authors
Jin, Y., Martinez‐Maldonado, R., Gašević, D., Han, X., & Yan, L.
Affiliations
  • Monash University (AU)
  • Faculty of Information Technology, Monash University Melbourne Australia
  • Tsinghua University (CN)
Corresponding author
  • Gašević, Dragan
  • Yan, Lixiang
Venue
Journal of Computer Assisted Learning
Publisher
Wiley
Published
2026-08-22
Volume
42
Issue
5
ISSN
  • 0266-4909 (ISSN-L)
  • 1365-2729
OpenAlex
W7204005753
Cited by
1
References
99
Citations by year
2026: 1
FWCI
1.08
Type
journal article
Language
en
Open access
Locations
  • Journal of Computer Assisted Learning (publishedVersion, cc-by-nc)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • HAL (submittedVersion)
  • arXiv (Cornell University) (cc-by)
Funders
  • National Natural Science Foundation of China (20261710003)
  • Australian Research Council (DP220101209, DP240100069)
  • Jacobs Foundation
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Intelligent Tutoring Systems and Adaptive Learning
  • Social Sciences › Psychology › Developmental and Educational Psychology › Innovative Teaching and Learning Methods
  • Social Sciences › Education › Artificial Intelligence in Education
Keywords
agentic AI, learner agency, human-AI collaboration, AI personas, co-creation, group dynamics, psychological safety, cognitive load, creative performance, discourse patterns, sequential pattern mining, Gaussian mixture model clustering

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-18, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-18 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.