Artificial intelligence in service-learning: A systematic mapping review and future research agenda for higher education

Synthesis: Yusuf, Atan and Ting (2026) ran a PRISMA-guided systematic mapping review of 26 peer-reviewed studies of AI in service-learning in higher education, retrieved from 745 records, and found that AI use clusters in preparation (18 studies) and action (15), thins at evaluation (8) and again at Reflection (3), while direct service (18) delivered onsite (13) remains the dominant configuration. The mapping's most consequential result is negative: the outcomes it records are the ones conventional service-learning already produces — knowledge application in 23 studies, transferable skills in 20, civic orientation in 14 — and only two studies report anything for community beneficiaries, so the corpus cannot separate what AI adds from what the pedagogy was already doing.

Key Findings

• The screen was tight and the residue small. Six databases plus Google Scholar produced 745 records; 52 duplicates left 693 for title-and-abstract screening, 83 went to full text, and 26 met every criterion. Two reviewers screened independently at both stages (Cohen's κ = 0.88 and 0.89), and coding agreement was κ = 0.91.
• The literature is two years old, not sixteen. The earliest included study is from 2010, but nothing else qualified until 2021. Ten studies appeared in 2024 and ten in 2025, together 76.9% of the corpus, plus one from 2026. Counting the appendix, the 26 studies sit in 11 settings: Hong Kong SAR, China (7), the United States (6), mainland China (3), Italy (2), and single studies from Spain, Mexico, South Korea, Colombia, Austria, Singapore and Peru; one does not state a country.
• AI supports the work around service more than the service itself. Direct service appears in 18 of 26 studies, indirect in six, research-based in three and advocacy in two, with five too thinly described to classify; onsite delivery appears in 13, hybrid in three and fully online in two, with eight unstated. The functions repeat the pattern: data analysis and interpretation (13), content generation and co-creation (10), adaptive support and personalization (8), coordination and decision support (5), tutoring and scaffolding (4), assessment and feedback (3), profiling and prediction (3). Two studies were coded as no function because AI was the subject taught, not a system doing work.
• Integration follows the phase, not the pedagogy. Preparation is the principal locus (18 studies), action follows (15), evaluation trails (8) and reflection is rare (3). In 11 of the 18 preparation studies, AI-supported groundwork carried into the action phase; only one study integrated AI across all four phases.
• Outcomes are student outcomes. Knowledge application (23), transferable skills (20), civic orientation (14), personal development (12), engagement (9) and motivation (5) dominate. Beneficiary outcomes appear in two studies. Eleven of the 26 reported at least one concern, unreliable output the most common (6), alongside readiness, over-reliance, contextual fit, ethical safeguards, and access and inclusion.
• Theory organizes the work but never explains it. Twelve of 26 studies drew substantively on a named theory or framework, yielding 20 applications: learning and pedagogy (6), design and process (5), knowledge and professional practice (4), instructional models (2), cognitive taxonomies (2), social and critical theory (1). Sixteen of the 20 informed design and targeting; none informed selection or tailoring, and none was tested or refined.

A small, concentrated corpus

Twenty-six studies is a thin base for a field the abstract says AI has begun to transform: service-learning, an experiential pedagogy. Of 610 records excluded at title and abstract, 517 did not pertain to service-learning at all; of 57 excluded at full text, 33 were not situated in an AI context. The exclusions say the two literatures barely touch: most candidates were either service-learning without AI or AI without service.

The timing matters more than the count. A corpus in which 76.9% of studies appeared in 2024 and 2025 is a research front, not a settled literature. The geography sharpens the point. Seven studies come from Hong Kong SAR, China and six from the United States; mainland China supplies three. A mapping of "higher education" therefore reports what a small number of institutions in three systems did, and the review has no cases from elsewhere with which to test whether community partnerships, credit structures or AI access change the pattern.

Why AI clusters in preparation and action

The distribution across phases is the review's strongest result because it records pedagogical choices, not tools. Preparation and action involve searching, generating, organizing and translating — tasks an AI system can absorb. Reflection requires interpretation, self-questioning and judgment, and the authors argue that current use follows "the path of least pedagogical resistance." Evaluation shows the same logic from the other side: six of the eight evaluation studies integrated AI only after the service was over, analyzing field diaries or 802 written reflections retrospectively rather than participating in the loop.

That is a problem for service-learning specifically, because reflection is the mechanism by which the service becomes learning. The corpus supports the operational half of the pedagogy and skips the interpretive half. It also contradicts what the reflection literature outside service-learning reports: GenAI reflection partners can improve written reflection quality and sustain journaling routines under the right conditions. The efficacy evidence exists; adoption has not followed it into community-engaged courses. The review cannot say whether reflection is harder to automate, distrusted in the interpretive space, or simply unattempted in early experiments — the coding records where AI was used, not why it was withheld.

What the outcomes cannot show, and what the theory cannot explain

The outcome list is a familiar one. Knowledge application, transferable skills, civic orientation and personal development are what meta-analyses of conventional service-learning already report, which is why the authors call attribution the field's central problem: a study reporting gains after an AI-supported service project has not shown that AI produced them, and few designs here could. The one directional signal runs the other way — in Kimmel (2024), a team abandoned ChatGPT over concerns that it would interfere with course objectives, and students who used it recorded lower examination scores — but it is a single course, reported as a caution rather than an effect estimate.

The beneficiary column is emptier still. Reciprocity is definitional to service-learning, yet only two studies report outcomes for the people served. The civic dimension is measured, if at all, on the student side: civic orientation appears in 14 studies as a student outcome, and the community's gain appears twice. A field that cannot show what partners received has not demonstrated the reciprocity it claims.

Theory is where the explanatory gap becomes a pattern. Twelve studies used a named framework, and 16 of 20 applications put it to work in design and targeting — structuring a module, sequencing a project, choosing content. Nine interpreted findings and six defined measurement. None tested a proposition or refined a theory, and selection and tailoring had no applications at all. The dominant families are learning and pedagogy and design and process; social and critical theory appears once. So the concerns the corpus documents — human judgment, ethical safeguards, relational fit — are precisely the ones its theories are least equipped to explain, which is the argument for theory development here rather than more design templates.

What this means for practice

• Instructors: design the reflection before the tool. If AI is used to plan and produce service materials, keep at least one reflective artifact that AI did not draft, and assess the interpretation rather than the artifact. The phase counts (preparation 18, action 15, reflection 3) describe the field's default; a course can choose differently.
• Instructional designers: specify a beneficiary outcome and collect it. With beneficiary outcomes in two of 26 studies, the community side of a partnership is usually unevidenced. Name one partner-level result — access, usability, a delivered resource — and measure it with the partner, not for them.
• Administrators: fund readiness and access, not licenses. Eleven of 26 studies reported concerns; unreliable output was the most common, and access barriers fell hardest on under-resourced partners. Budget for verification time, training and devices, and treat equity as a deployment condition.
• Instructors: keep the human in the assessment loop. Where AI gave feedback (three studies), the strongest example kept teachers responsible for final grades and able to adjust AI judgments. Formative use is what the corpus supports; automated judgment is not.
• Researchers: comparison before enthusiasm. The field needs designs that separate AI's contribution from service-learning's established effects, measures shared across studies, and theory that is tested rather than applied.

Limitations

The authors name two limits: no critical quality appraisal, defended as appropriate to a mapping review's job of classification rather than evidence grading, and a corpus of 26 studies. The first still matters — the frequency counts treat a single-course case study and a 794-student action research project as equal units, and several studies report no sample size.

Search strings were adapted per database, so recall differs by index; only English and Chinese publications were eligible; the 2026 sweep closed before the year ended. Eight studies had no clear delivery configuration and five no clear service form, so the configuration percentages rest on the classifiable subset. The AI boundary also leaks: two included studies were coded as no function because AI was the content taught, so the corpus mixes AI-as-tool with AI-as-subject. No study has a control group or a longitudinal follow-up, so every outcome statement reports what authors observed, not an effect. The research agenda is also a proposal, not a finding: its gaps are inferred from the counts above.

Citation

Yusuf, A., Atan, N. A., & Ting, I. T. Y. (2026). Artificial intelligence in service-learning: A systematic mapping review and future research agenda for higher education. Review of Education, 14, e70251.