Generative AI and the Community of Inquiry Framework: A Scoping Review of Teaching, Social, and Cognitive Presence

Synthesis: Doo and Park (2026) report a PRISMA-guided scoping review of 21 empirical studies published between November 2022 and July 2026 on how generative AI relates to teaching, social, and cognitive presence in the Community of Inquiry framework. The map is uneven. Social presence appears in 13 of the 21 studies (61.9%), cognitive presence in 10 (47.6%), and teaching presence in only 5 (23.8%), and no study examined teaching presence on its own. Cognitive presence was most often associated with exploration and integration while resolution remained the least evident phase, social presence responded to anthropomorphic design cues, and teaching presence surfaced mainly when AI was assigned an explicit instructional role. The authors read generative AI as a design-sensitive pedagogical mediator rather than a uniform support for CoI, and they pool no effects across studies.

Key Findings

1. Twenty-one studies, mapped rather than pooled. The review retained 21 peer-reviewed empirical studies published from November 2022 through July 2026, drawn from 16 journals and 8 countries, with 7 studies added during manuscript revision.
2. The evidence base is concentrated. The United States supplied 8 studies (38.1%) and China 6 (28.6%), 16 of the 21 studies (76.2%) involved undergraduates, and ChatGPT was the named tool in 7 studies (33.3%) while chatbots accounted for 4 (19.0%).
3. Cognitive presence clusters in the middle phases. Generative AI was most consistently associated with exploration and integration, and resolution was the least evident phase across the studies that coded all four phases of the practical inquiry model.
4. Research attention is structurally uneven. Social presence was examined in 13 of 21 studies (61.9%), cognitive presence in 10 (47.6%), and teaching presence in 5 (23.8%); every teaching-presence study examined it alongside another presence.
5. Teaching presence appears when AI is given an instructional role. It was strengthened when AI acted as a tutor, as a co-facilitator beside a teacher, or as the source of learning analytics, and it did not move when only the type of AI Feedback changed.
6. Instructional configuration sorts the outcomes. Discussion-based and task-based designs were most often associated with cognitive presence, embodied AI instructor designs with social presence, and self-regulated learning designs with social presence acting as a pathway to learner autonomy.

How the review was assembled

The authors followed PRISMA guidelines and searched Google Scholar, Education Source, ERIC, PsycINFO, and Web of Science, supplementing the database search with manual scans of leading educational technology and online learning journals. Studies qualified if they were empirical, published in English between November 2022 and July 2026, and reported quantitative, qualitative, or mixed-method data on at least one CoI presence. Conceptual papers, reviews, and meta-analyses were excluded, so the 21 studies are primary evidence rather than secondary summaries.

Coding covered publication details, learner level, subject area, instructional approach, type of generative AI, and presence dimension. Two authors coded independently with high agreement (Cohen's k = .82), resolved disagreements by discussion, and appraised each study with the JBI checklist matching its design. No major methodological concerns were identified. Synthesis was narrative, organized by the three presences and by the process mechanisms linking AI use to learning, which is why the paper reports recurring patterns rather than an averaged effect.

Publication was concentrated: three studies appeared in 2022, two in 2023, one in 2024, seven in 2025, and eight in 2026. The 21 studies spanned 8 countries and 16 journals, with the British Journal of Educational Technology contributing 4 (19.0%). ChatGPT was the named tool in 7 studies (33.3%), chatbots in 4 (19.0%), and AI-generated instructors or pedagogical agents in 2 (9.5%); the remaining 8 (38.1%) used BERT, SPARK, embedded analytics, or other generative systems.

An uneven map of three presences

The distribution is easy to misread. Social presence being the most frequently examined dimension does not mean it has the strongest evidence; it means more studies asked about it. Eight studies focused on social presence alone and six on cognitive presence alone, while teaching presence never appeared as a standalone construct. All five teaching-presence studies examined it together with cognitive or social presence, which the authors treat as evidence that teaching presence has not yet been independently theorized in the generative AI literature.

The imbalance runs against the framework's own logic. Teaching presence is the integrative force that orchestrates cognitive and social processes, so its relative absence leaves the field describing effects on inquiry without describing who designs and steers it. When teaching presence did appear, it was usually a by-product of a broader design choice: an AI tutor supplying personalized guidance, an AI co-facilitator beside a human teacher in micro-blended mathematics, or analytics embedded in a platform. In one marketing course comparing AI as tutor, tool, and teammate, participants rated cognitive presence effective far more often (85.71%) than teaching presence (42.86%) or social presence (35.71%), a single-course result reported as illustration rather than estimate.

Measurement compounds the unevenness. Some studies treated presence as a perceived outcome on self-report scales, others inferred it from discourse patterns, interaction logs, or design features, and still others coded specific phases of the practical inquiry model. Self-report social presence and discourse-coded cognitive presence are not interchangeable quantities, so cross-study comparison is limited even where the same presence is named.

What different instructional designs activate

The review groups the studies into four instructional configurations, and the grouping carries most of its practical value. In online discussion environments, generative AI scaffolded inquiry, structured argumentation, and idea elaboration; these designs were most consistently associated with cognitive presence, particularly exploration and integration. Idea-oriented feedback in inquiry-based discussion raised cognitive presence relative to task-oriented feedback, while task-oriented feedback was the condition that strengthened social presence through open communication.

In AI-generated instructor and video-based settings, the relevant design variables were behavioral fidelity, voice realism, dynamic visual cues, and embodied tutor behavior. These designs were most often associated with social presence. Studies in this group reported higher perceived social presence, satisfaction, and motivation when the agent behaved more like a person, and one study found that identity anthropomorphism raised social presence through positive emotion.

Self-regulated learning designs used chatbots for goal setting, help-seeking, self-evaluation, and monitoring in MOOCs and fully online courses. Here social presence frequently operated as a mediating mechanism linking chatbot interaction to learner autonomy, which makes this the clearest strand in the review where a presence is treated as a pathway rather than an outcome.

Task-specific collaborative activities, including AI-assisted writing, lesson planning, conversation practice, and analytics-supported collaborative learning, placed AI as a teammate or cognitive tool. These designs contributed most strongly to cognitive presence while also shaping social engagement during dialogue. Across all four groups, teaching presence was most evident when the AI held an explicit instructional role rather than acting as a general-purpose assistant.

Where presence did not translate into learning

The review is careful about null and negative results. In one primary-school mathematics study, an animated pedagogical agent with high behavioral fidelity raised social presence, satisfaction, and motivation and reduced cognitive load, yet did not significantly affect mathematics achievement either directly or as a mediator. In another, an identity-anthropomorphized virtual companion improved learning outcomes but students in that condition spent significantly less time on the instructional videos, which the authors attributed to the companion drawing attention away from the lesson.

Feedback design produced a similar trade-off. Learners who received idea-oriented prompts showed higher cognitive engagement than those who received task-oriented prompts but also reported higher negative emotional engagement. The effect of analytics-supported learning was contingent on the level of CoI support and on students' digital competencies. And in a longitudinal study of chatbot interaction, more frequent interaction was associated with decreased social presence and lower autonomy for students oriented toward knowledge acquisition, even though social presence fully mediated the autonomy benefit for students who preferred virtual companionship.

These results support the review's central caution: generative AI selectively activates particular presences depending on instructional format, AI role, and design features, not all three at once. Few studies tested presences as mediators, so the mechanism evidence remains thin.

What this means for practice

• Instructors. Choose the AI role from the presence you want to move, not from the tool's feature list: assign it tutor duties to support teaching presence, teammate or cognitive-tool duties inside discussion and task work to support cognitive presence, and embodied instructor duties when social presence is the target.
• Instructors. Do not stop at exploration. The reviewed studies repeatedly found idea generation and integration without resolution, so pair AI-supported discussion with a deliberate closing task that requires learners to apply, test, or defend a conclusion.
• Instructional designers. Build the social and regulatory supports around the AI rather than assuming they come with it. Goal setting, help-seeking, and self-evaluation prompts were the designs in which social presence worked as a pathway to autonomy, and the anthropomorphism studies show that stronger human-likeness can also pull attention away from content.
• Instructional designers. Treat teaching presence as a design object when AI shares instruction with a human. The review found teaching presence strongest when AI co-facilitated beside a teacher or supplied analytics inside a structured course, so specify the division of instructional labor rather than leaving it to the model.
• Researchers. Report how presence was operationalized, because self-report, discourse coding, and design-feature inference produced the same labels for different constructs; teaching presence needs studies that isolate it rather than measuring it as a residual of multi-presence surveys.

Limitations

• The review rests on 21 empirical studies, all English-language and all published after November 2022, so non-English work and earlier AI-in-education research are outside its scope and the distribution across presence dimensions is uneven.
• Presence definitions and operationalizations varied across the included studies, from self-report scales to discourse coding to inferred design characteristics, which limits direct comparison and makes the findings patterns rather than stable regularities.
• Learning presence, the proposed fourth element covering self-, co-, and shared regulation, was used as an interpretive lens but not as a primary analytical category, and its conceptual status remains debated.
• The term "presence" was used inconsistently in the source literature, including teacher presence, AI presence, and learner presence, and the review resolved this by adhering to Garrison's three-presence model, which may flatten distinctions other researchers consider meaningful.
• Generative AI changed rapidly during the review window, so many included implementations are early-stage, and no study tested longer-term effects on learner relationships or engagement.

Citation

Doo, M. Y., & Park, Y. (2026). Generative AI and the Community of Inquiry Framework: A Scoping Review of Teaching, Social, and Cognitive Presence. Online Learning, 30(SI2), e5880.