Systematic Review of Collaborative Learning Activities for Promoting AI Literacy

Synthesis: Systematic review of collaborative learning for AI literacy — SEFI 2025. A theory-driven, PRISMA-guided systematic review of 9 studies (2015–2023) examining how collaborative learning (CL) approaches can be harnessed to build AI Literacy. Using the ICAP framework (Interactive–Constructive–Active–Passive) as an analytical lens, the review finds that CL effectively increases AI literacy across a range of activities, settings, and learner groups, and that successful interventions engaged learners at multiple ICAP modes.

Background: why collaborative learning for AI literacy

AI literacy refers to the knowledge and skills necessary to understand, interact with, and make value-informed decisions about AI. As AI adoption accelerates across academia and industry, educators have designed instructional activities to make complex AI concepts accessible to learners of all ages. Prior work has unequivocally shown that group- and team-based collaborative learning (CL) benefits STEM, computing, and engineering education, yet little attention had been paid to how the interaction modality affects AI learning. This review uses a theory-driven approach to ask whether the well-established benefits of CL for information literacy extend to AI literacy, and how the mode of interaction shapes learning outcomes.

Theory guiding the review: the ICAP framework

The authors leveraged the ICAP framework (Chi & Wylie, 2014) to systematically compare efficacy across heterogeneous studies. ICAP distinguishes observable states of learning along a hierarchy of four modes: interactive (learners co-construct new knowledge by asking questions, giving and receiving feedback, and incorporating others' ideas), constructive (learners go beyond the initial instructions to generate new output), active (knowledge is applied to similar but non-identical scenarios), and passive (listening or reading, which is least effective because it requires similar cues or contexts for recall). Interactive activities that co-construct new knowledge yield the best learning outcomes. ICAP was chosen because it is an effective lens for evaluating CL in CS and engineering education and provides stable markers for comparing studies.

Review methodology

The review followed the PRISMA methodology. A February 2024 search across ACM Digital Library, IEEE Xplore, SCOPUS, and Web of Science yielded 227 candidate studies (ScienceDirect and ProQuest were excluded for excessive irrelevant results). After pre-screening for duplicates and retractions, two reviewers independently applied six inclusion criteria — published 2015–2023, conference/journal article, in English, focused on AI literacy, involving human–human or human–AI collaboration, and presenting an empirical activity or intervention. Inter-rater agreement was high (Cohen's κ = 0.90), and discrepancies were resolved by discussion. The most common reason for exclusion was failing to describe an implemented activity in sufficient detail to assess its effectiveness. This process yielded nine included studies.

Key Findings

1. Collaborative learning effectively raises AI literacy. Across the corpus, CL was found to increase AI literacy across a wide range of activities, settings, and learner populations — confirming that the benefits of collaboration for information literacy generalize to the AI domain.
2. Successful interventions engaged learners at multiple ICAP modes. All nine studies included multiple ICAP modes, typically combining more individualized learning (passive/active) with collaborative components. Across the corpus: 9 studies included passive components, 5 active, 7 constructive, and 8 interactive. Notably, no study relied on a single mode — the strongest designs combined lower and higher modes to fit learners' skill levels.
3. The Active→Constructive transition matters most. Following Chi & Boucher (2023), the authors emphasize that the most significant leap in learning occurs when moving from Active to Constructive modes — equipping learners to create new knowledge. The reviewed interventions were structured to enable this transition, moving most learners from introductory to intermediate contexts.
4. Collaboration spans humans and AI. Studies involved student-to-student, educator-to-educator, family-member-to-family-member, and human-with-AI partnerships. AI agents appeared as collaborative partners (e.g., students working with AI in Druga 2019 and Wan 2020), pointing to a growing role for human–AI collaboration in AI literacy.
5. Classroom-centric but broadening. Most studies occurred in classrooms, but participation was deliberately broadened in several — involving parents and families (Long et al. 2022), educators and facilitators (Lee et al. 2022), and AI partners — connecting to Co-Designing Community-Centered AI Education for Adults: A Midwestern Case Study and Programming Language Policy as an AI Literacy Equity Problem: A 15-Nation Comparative Analysis.
6. Limited but growing evidence base. Nine studies is a small corpus reflecting the nascency of empirical research at this intersection. The authors call for more rigorous designs, larger samples, and longitudinal tracking of AI literacy development.

What this means for practice

• Instructors. Incorporate collaborative learning activities into AI literacy instruction so learners co-create knowledge contextualized to their own needs.
• Instructors. Target the instructional goal when choosing an activity — literacy, specific use-case knowledge, or domain knowledge — and pick the ICAP mode that fits it.
• Instructors. Give learners opportunities to engage with foundational concepts before they synthesize and generate new ideas, with attention to the Active→Constructive transition that most advances comprehension.
• Researchers. Extend the evidence across diverse settings with more design-based research within a setting, to establish whether multiple iterations improve outcomes.
• Researchers. Investigate how AI partners can best support the learning of AI itself and how human–human versus human–AI interaction shapes learners’ understanding of AI’s capabilities and constraints, noting that the corpus is disproportionately US-based and that AI literacy need not take a singular global form but should be contextual to local communities and may be integrated with data or digital literacy.

Limitations

• The nine studies were implemented in heterogeneous learning contexts, which makes cross-study synthesis difficult.
• The search may have excluded relevant studies that used different terminology for the same constructs.
• The controlled learning environments described may omit elements of real-world collaboration — interactions with colleagues, repositories, or other resources — that are hard to replicate.

Citation

Hingle, A., & Johri, A. (2025). Systematic review of collaborative learning activities for promoting AI literacy.