Human–AI Collaboration Reconfigures Group Regulation from Socially Shared to Hybrid Co-Regulation ⋆
Yujing Zhang1 , Xianghui Meng1 , Shihui Feng1 , and Jionghao Lin1 ,2,3
arXiv:2604.08344v1 [cs.AI] 9 Apr 2026
1
The University of Hong Kong, Hong Kong SAR, China {u3011372, u3014272}@connect.hku.hk {shihuife, jionghao}@hku.hk 2 Carnegie Mellon University, Pittsburgh, PA, USA 3 Monash University, Clayton, VIC, Australia
Abstract. Generative AI (GenAI) is increasingly used in collaborative learning, yet its effects on how groups regulate collaboration remain unclear. Effective collaboration depends not only on what groups discuss, but on how they jointly manage goals, participation, strategy use, monitoring, and repair through co-regulation and socially shared regulation. We compared collaborative regulation between Human–AI and Human– Human groups in a parallel-group randomised experiment with 71 university students completing the same collaborative tasks with GenAI either available or unavailable. Focusing on human discourse, we used statistical analyses to examine differences in the distribution of collaborative regulation across regulatory modes, regulatory processes, and participatory focuses. Results showed that GenAI availability shifted regulation away from predominantly socially shared forms towards more hybrid co-regulatory forms, with selective increases in directive, obstacleoriented, and affective regulatory processes. Participatory-focus distributions, however, were broadly similar across conditions. These findings suggest that GenAI reshapes the distribution of regulatory responsibility in collaboration and offer implications for the human-centred design of AI-supported collaborative learning. Keywords: Generative AI · Computer-Supported Collaborative Learning (CSCL) · Socially Shared Regulation · Co-Regulation · Human–AI Collaboration
1
Introduction
Generative artificial intelligence (GenAI) in computer-supported collaborative learning (CSCL) is increasingly conceptualised not only as a tool, but also as an interactive contributor that can generate ideas, challenge assumptions, and extend collective reasoning during group work [14]. However, effective collaboration depends not simply on participation, but on how joint activity is organised ⋆
Corresponding author.
2
Y. Zhang et al.
through shared planning, monitoring, evaluation, and adaptation over time [12]. In collaborative contexts, regulation extends beyond individual self-regulated learning to include co-regulation (CoRL) and socially shared regulation (SSRL), both of which are central to effective collaboration because they rely on coordination and shared responsibility among group members [6]. Despite their importance, empirical evidence on how GenAI relates to CoRL and SSRL remains limited [9]. This gap is increasingly important as GenAI is integrated into collaborative learning settings without closely examining group regulation. Recent studies suggest that GenAI can reshape participation and interaction in collaborative problem solving [5], raising concerns about how regulatory responsibility is redistributed when GenAI enters group work. Addressing this gap is therefore important for advancing Human–AI collaboration theory and informing human-centred AI design in education. To examine whether GenAI availability reconfigures collaborative regulation, this study designed and deployed a conversational GenAI agent explicitly oriented towards supporting collaborative regulation. Rather than providing direct solutions, the agent intervened through dialogue to shape how groups coordinated and regulated their joint activity. By comparing Human–AI and Human– Human groups working on the same tasks, we examined differences in the distribution of collaborative regulation across regulatory modes, regulatory processes, and participatory focuses. Specifically, we asked: What differences are observed in the distribution of collaborative regulation between Human–AI and Human– Human groups?
2
Related Work
2.1
Co-Regulation and Socially Shared Regulation in CSCL
When collaborative tasks become complex and open-ended, groups must regulate their activity through co-regulated learning (CoRL), and socially shared regulation of learning (SSRL) [6]. CoRL and SSRL are interactional and enacted through dialogue. CoRL describes peer-supported regulation (e.g., prompts, guidance, feedback) that helps individuals manage their learning, while SSRL involves the group’s deliberate, collective regulation of goals, strategies, monitoring, and reflection. CoRL often serves as a bridge that enables SSRL to emerge and stabilise [6]. As such, SSRL is widely treated as a hallmark of effective collaboration, capturing moments when groups jointly frame the task, negotiate approaches, track progress, and adapt to challenges [6,7]. In CSCL, however, these forms of regulation rarely arise reliably without support. Learners bring differing goals, knowledge, and affective states, so simply placing them together does not ensure the development of CoRL or SSRL [10]. Productive collaboration requires ongoing alignment, including shared task understanding, role coordination, progress monitoring, and expectation management, and these regulatory demands shift as interaction unfolds [8]. This makes scaffolding CoRL and SSRL a central design problem in CSCL and a key target for interventions intended to improve collaborative process and outcomes [3].
Human–AI Collaboration Reconfigures Group Regulation
2.2
3
GenAI Agent in CSCL
Recent advances in GenAI have opened new possibilities for CSCL interventions: beyond information delivery, GenAI systems can generate ideas, pose questions, and scaffold reasoning, suggesting potential roles in shaping collaborative learning processes [1]. Yet, despite the central role of co-regulation and socially shared regulation in effective CSCL, few studies have examined whether GenAI can meaningfully shape collaborative regulation in group learning. A notable early study introduced a GenAI agent to prompt metacognitive reflection in support of SSRL, but found limited evidence of improved regulation when assessed through linguistic alignment [3]. Crucially, we argue that this reflects an analytic limitation rather than a failure of AI support: treating regulation as a static shift in language use is insufficient to capture its distributed and emergent nature. Building on this insight, we treat CoRL and SSRL as interactional phenomena that can be observed through the distribution of regulatory modes, regulatory processes, and participatory discourse moves in group interaction.
3
Method
3.1
Data Collection
Figure 1 summarises the data collection procedure. We used a between-groups experimental design: participants were randomly assigned to a GenAI-available or GenAI-unavailable condition and then formed into small groups within condition to complete the same collaborative tasks. The sample comprised 71 university students (aged 18–35; M = 23.73, SD = 3.65) from English-speaking countries and was predominantly Asian (n = 66; mixed ethnicity n = 2; White n = 3). Participants were allocated to 24 groups (23 triads and one dyad). The study received ethical approval from the Human Research Ethics Committee of The University of Hong Kong (Ref. No.: EA250843).
Human-AI Condition (12 Groups)
Randomised Pre Registration Survey
Follow-up Survey
Participants N = 71
Human-Human Condition (12 Groups)
Fig. 1. Study Procedure.
3.2
Group Task Design
The group tasks were grounded in the collective intelligence paradigm, which posits that group performance reflects a stable collective intelligence factor rather
4
Y. Zhang et al.
than task-specific skills alone [13]. To capture different coordination and perspectivetaking demands, we used two complementary tasks: a moral-reasoning case on academic misconduct that required value-based discussion and justified consensus among multiple-choice options, and a shopping-planning task that required integrating constraints and producing a concrete route plan. 3.3
GenAI Agent Design
The GenAI agent (OpenAI GPT-4o mini) was embedded in an online CSCL platform to support real-time, text-based group collaboration (see Fig.2). Table 1 summarises the conversational prompts used by the GenAI agent to support collaborative regulation during group work. Rather than providing content or solutions, these prompts were designed to elicit explanation, encourage perspective-taking, and guide groups through progression of their joint activity.
Call Moderator
Request Output
Leave Room Audio Output
You have joined Group 27 as Aisha. The topic is Global energy production.
Audio Input
Jacob has joined.
MODERATOR
Hello everyone! Welcome to the discussion on the future of energy production. Each of you will present on one of the energy forms: wind, solar, or nuclear. Make sure to cover both the advantages and disadvantages of your assigned energy form. Let‘s start with a brief introduction from each of you about which energy source you will be discussing. Feel free to share your initial thoughts as well. Aisha, would you like to begin? 06/02/25, 09:34
Aisha
Hi l‘m Aisha, currently pursuing my Masters in Enviormental Science. l’m particularly interested in how we can harness the power of renewable sources like solar and wind to achieve sustainable energy solutions. l‘ve seen Evidence. Especially in India and Denmark. Showing how decentralized power grids improve resilience and reduce carbon emissions in the long run. 06/02/25, 09:34
Enter message
Send
Fig. 2. GenAI agent interface embedded in the platform.
3.4
Measure
Group Regulatory Dynamics Group regulatory dynamics were analysed using a two-level coding framework that distinguishes extended regulatory activity from moment-to-moment discourse actions. At the episode level, a regulatory episode was defined as a continuous segment focused on a single regulatory issue, typically spanning multiple turns. Episodes
Human–AI Collaboration Reconfigures Group Regulation
5
Table 1. GenAI Prompts Used to Support Collaborative Regulation Intervention Behaviour
Example Prompt
Further Reasoning
“Could you explain your answer in more detail?”
Invite Examples
“Can you give an example to support your point?”
Challenge Viewpoints
“Do you think this always holds, or might there be exceptions?”
Turn-Taking Management
“Thank you Jack! Anyone else want to share their view?”
Summarising
“Let me summarise the group discussion.”
Task Transition
“How would you like to move forward from here?”
Invite Building on Ideas
“How does this connect with what was mentioned earlier?”
began when a regulatory concern was introduced and ended when it was resolved or the group shifted focus. Each episode was coded for regulatory mode: CoRL, SSRL, or hybrid CoRL&SSRL (i.e., individual guidance and shared regulation co-occurring within the same episode). Episodes were then coded for specific regulatory processes (adapted from Nguyen et al.[11]; see Table 2). At the utterance level, each conversational turn was coded for participatory focus (adapted from Dang et al.[2]; see Table 3). This captures the speaker’s immediate regulatory action(e.g., planning next steps, reasoning about options, reporting outcomes) [4]. Table 2. Episode-Level Regulatory Processes Code
Subcode Obstacle Detection
CoRL
Strategic Action Direction Affective Support
Operational definition Identifying an obstacle in another member’s learning or task progress. Directing another member to adopt or change a strategy. Regulating another member’s emotional or motivational state.
Shared Obstacle Negotiation
Jointly identifying and discussing group-level obstacles.
Shared Strategic Negotiation
Jointly proposing and negotiating alternative strategies.
Shared Action Change
Jointly agreeing on a change of action and coordinating execution.
Shared Affective Regulation
Jointly regulating the group’s emotional or motivational state.
SSRL
Note. A and B denote different human group members.
Example “You seem stuck on this step.” “Jack, you can use the formula instead.” “Don’t worry, Lily! ” A: “We’re off track.” B: “Yeah, we haven’t answered the question.” A: “Can we compare both cases? ” B: “Yeah, let’s start with the first one.” A: “Let’s change our method.” B: “Okay, let’s do that.” A: “This is frustrating.” B: “Yeah, but we can calm down and keep going.”
6
Y. Zhang et al. Table 3. Utterance-Level Participatory Focus
Code
Subcode Cog–Explain
Cognitive Cog–Reason Cog–Evaluate Meta–Orient Metacognitive
Meta–Plan Meta–Monitor Meta–Reflect
Task execution
TE–Act TE–Report
Socio-emotional
SE–Express SE–Regulate
Operational definition Explains task concepts or rules to build shared understanding. Uses reasoning to justify choices or compare options. Judges the quality or fairness of an option. Frames what the group should decide or focus on. Proposes strategies, sequencing, or next steps. Checks progress, time, understanding, or agreement. Evaluates past strategies or collaboration quality. Indicates execution of a concrete task action. Reports intermediate or final task results.
Example “Frozen food must be returned within 40 minutes.” “If we exceed £50, we lose points.” “Choice A is too harsh.” “Let’s start with Question 1.” “Everyone give your answer, then we compare.” “How much time do we have left? ” “This approach isn’t working well.” “I’ll calculate this part.” “The answer is 175.”
Expresses an emotional “This is funny lol.” reaction. Regulates group emotion or “Good job guys! ” motivation.
Two trained coders independently annotated the data. Inter-rater agreement was near perfect for utterance-level participatory focus (κ = .85–.98) and episode-level regulatory processes (κ = .86–.95), with high span-based episode segmentation accuracy (macro F1 = .94, boundary F1 = .97). This span-level evaluation was adopted because regulatory episodes vary in length, and reliable coding requires both consistent category assignment and accurate identification of episode boundaries.
4
Data Analysis
All analyses were conducted at the group level; in the Human–AI condition, GenAI-generated turns were excluded so that findings reflected how GenAI availability reshaped human regulatory organisation rather than effects driven by GenAI utterance volume or content. We used mixed-design ANOVAs because category (i.e., regulatory mode, regulatory process, or participatory focus) was a within-group factor measured repeatedly, whereas condition (i.e., Human–AI vs Human–Human) was a betweengroup factor. Post hoc Holm-adjusted pairwise contrasts were then used to compare conditions within each category while controlling the family-wise error rate across multiple tests.
Human–AI Collaboration Reconfigures Group Regulation
5
Results
5.1
Human–AI Groups Shifted Towards Hybrid Co-regulatory Organisation
7
Episode-level regulation mode. A mixed-design ANOVA on regulatorymode proportions revealed a significant main effect of mode, F (1.20, 26.31) = 89.28, p < .001, indicating that regulatory responsibility was unevenly distributed across modes. There was no main effect of condition, F (1, 22) ≈ 0.00, p > .999, but a significant Condition × Mode interaction, F (1.20, 26.31) = 222.46, p < .001, indicating that regulatory organisation differed between conditions. Holmadjusted contrasts (see Table 4) showed that Human–Human groups exhibited a higher proportion of SSRL episodes, whereas Human–AI groups showed higher proportions of hybrid CoRL&SSRL and CoRL-only episodes. Overall, GenAI availability was associated with a shift from predominantly socially shared regulation towards more hybrid and co-regulatory organisation. Table 4. Distribution of regulation modes by condition Regulation mode CoRL SSRL CoRL&SSRL
Human–AI
Human–Human
Mean Prop. (SD) 0.066 (0.08) 0.083 (0.11) 0.851 (0.16)
Mean Prop. (SD) 0.000 (0.00) 0.940 (0.12) 0.060 (0.12)
t statistic (df = 22) −2.87∗∗ 18.29∗∗∗ −13.54∗∗∗
Note. Mean Prop. indicates the average proportion of a group’s total regulatory episodes accounted for by each mode. ∗∗∗ p < .001; ∗∗ p < .01.
Episode-level regulation process. A mixed-design ANOVA on regulatoryprocess proportions revealed significant main effects of condition, F (1, 22) = 98.07, p < .001, and process category, F (3.56, 78.23) = 63.88, p < .001, as well as a significant Condition × Process interaction, F (3.56, 78.23) = 33.53, p < .001, indicating process-specific condition differences. Holm-adjusted contrasts (Table 5) showed that Human–AI groups had higher proportions of Affective Support, Obstacle Detection, Shared Affective Regulation, and Strategic Action Direction. Overall, GenAI availability was associated with a selective reconfiguration of regulatory processes rather than a uniform shift across functions. Utterance-level participatory focus. A mixed-design ANOVA on participatoryfocus proportions showed a significant main effect of focus category, F (3.56, 78.41) = 92.72, p < .001, and a small main effect of condition, F (1, 22) = 4.70, p = .041, but no Condition × Focus interaction, F (3.56, 78.41) = 1.44, p = .233. Holmadjusted contrasts revealed no reliable condition differences within any focus category (all p ≥ .0646). This pattern suggests that although overall participatory focus distributions differed slightly between conditions, GenAI availability did not selectively shift attention toward any specific focus type.
8
Y. Zhang et al. Table 5. Distribution of Episode-Level Regulatory Processes by Condition
Regulatory process Affective Support Obstacle Detection Strategic Action Direction Shared Obstacle Negotiation Shared Strategic Negotiation Shared Action Change Shared Affective Regulation
Human–AI
Human–Human
Mean Prop. (SD) 0.837 (0.20) 0.314 (0.21) 0.808 (0.21) 0.284 (0.17) 0.829 (0.18) 0.817 (0.20) 0.838 (0.22)
Mean Prop. (SD) 0.00 (0.00) 0.183 (0.21) 0.060 (0.12) 0.394 (0.19) 0.824 (0.17) 0.778 (0.22) 0.128 (0.18)
t statistic (df = 22) −14.62∗∗∗ −2.20∗ −10.58∗∗∗ 1.26 −0.07 −0.46 −9.14∗∗∗
Note. Mean Prop. indicates the average proportion of a group’s total regulatory episodes accounted for by each process category. ∗∗∗ p < .001; ∗ p < .05.
6
Discussion and Conclusion
This study provides empirical evidence that GenAI availability is associated with changes in the distribution of collaborative regulation in CSCL. Compared with Human–Human groups, Human–AI groups showed less predominantly socially shared regulation and more hybrid co-regulatory forms, together with higher proportions of directive, obstacle-oriented, and affective regulatory processes. Participatory-focus distributions, however, were broadly similar across conditions, suggesting that GenAI was associated more with a redistribution of regulatory responsibility than with broad changes in moment-to-moment discourse. These findings suggest that GenAI may reconfigure collaborative regulation by changing how groups coordinate support, monitoring, and shared control. For practice, they imply that educational GenAI should be designed to scaffold rather than replace shared regulation during collaboration. This study has several limitations. The sample was relatively small and demographically skewed, many participants used English as a second language, and the study focused on text-only interaction. Future research should examine whether similar patterns emerge in more diverse and multimodal collaborative settings.
Acknowledgment This work was supported by the Faculty Research Fund and by the grant from the URC (Grant No. 2401102970) at The University of Hong Kong.
References 1. Bansal, G., Chamola, V., Hussain, A., Guizani, M., Niyato, D.: Transforming conversations with ai—a comprehensive study of chatgpt. Cognitive Computation 16(5), 2487–2510 (2024)
Human–AI Collaboration Reconfigures Group Regulation
9
2. Dang, B., Nguyen, A., Järvelä, S.: Deliberative interactions for socially shared regulation in collaborative learning. Journal of Learning Analytics 11(3), 192–209 (12 2024). https://doi.org/10.18608/jla.2024.8393 3. Edwards, J., Nguyen, A., Lämsä, J., Sobocinski, M., Whitehead, R., Dang, B., Roberts, A.S., Järvelä, S.: Human-ai collaboration: Designing artificial agents to facilitate socially shared regulation among learners. British Journal of Educational Technology 56(2), 712–733 (2025) 4. Feng, S., Wei, Z.: The CoMPAS framework for modeling individual participation in collaborative learning processes: a systematic review. International Journal of Computer-Supported Collaborative Learning 20, 425–461 (2025). https://doi. org/10.1007/s11412-025-09450-x 5. Feng, S.: Group interaction patterns in generative ai-supported collaborative problem solving: Network analysis of the interactions among students and a gai chatbot. British Journal of Educational Technology (2025) 6. Hadwin, A., Järvelä, S., Miller, M.: Self-regulation, co-regulation, and shared regulation in collaborative learning environments. In: Schunk, D.H., Greene, J.A. (eds.) Handbook of Self-Regulation of Learning and Performance, pp. 83–106. Routledge/Taylor & Francis Group, 2 edn. (2018). https://doi.org/10.4324/ 9781315697048-6 7. Järvenoja, H., Järvelä, S., Malmberg, J.: Understanding regulated learning in situative and contextual frameworks. Educational Psychologist 50(3), 204–219 (2015) 8. Järvelä, S., Kirschner, P.A., Hadwin, A., Järvenoja, H., Malmberg, J., Miller, M., Laru, J.: Socially shared regulation of learning in CSCL: understanding and prompting individual- and group-level shared regulatory activities. International Journal of Computer-Supported Collaborative Learning 11(3), 263–280 (8 2016). https://doi.org/10.1007/s11412-016-9238-2 9. Kim, J., Detrick, R., Yu, S., Song, Y., Bol, L., Li, N.: Socially shared regulation of learning and artificial intelligence: Opportunities to support socially shared regulation. Education and Information Technologies 30(9), 11483–11521 (1 2025). https://doi.org/10.1007/s10639-024-13187-9 10. Lyons, K.M., Lobczowski, N.G., Greene, J.A., Whitley, J., McLaughlin, J.E.: Using a design-based research approach to develop and study a web-based tool to support collaborative learning. Computers & Education 161, 104064 (10 2020). https: //doi.org/10.1016/j.compedu.2020.104064 11. Nguyen, A., Järvelä, S., Rosé, C., Järvenoja, H., Malmberg, J.: Examining socially shared regulation and shared physiological arousal events with multimodal learning analytics. British Journal of Educational Technology 54(1), 293–312 (10 2022). https://doi.org/10.1111/bjet.13280 12. Rogat, T.K., Linnenbrink-Garcia, L.: Socially shared regulation in collaborative groups: An analysis of the interplay between quality of social regulation and group processes. Cognition and Instruction 29(4), 375–415 (2011). https://doi.org/10. 1080/07370008.2011.607930 13. Woolley, A.W., Chabris, C.F., Pentland, A., Hashmi, N., Malone, T.W.: Evidence for a collective intelligence factor in the performance of human groups. Science 330(6004), 686–688 (2010). https://doi.org/10.1126/science.11931 14. Yan, L., Greiff, S., Teuber, Z., Gašević, D.: Promises and challenges of generative artificial intelligence for human learning. Nature Human Behaviour 8, 1839–1850 (2024)