Generative AI and student motivation: an SDT-guided systematic review of learning activities, motivational regulation, and measurement

Synthesis: Thirty-nine empirical studies, published 2023 to 2025 and drawn from Web of Science and Scopus, ask what generative AI does for student motivation. Read through Self-Determination Theory, the corpus is lopsided. Competence activities outnumber autonomy and relatedness ones, feedback and explanation dominate, and nobody tests whether an activity filed under a need actually satisfies it. Motivation is measured almost entirely by questionnaire, and motivational regulation, the part of the theory that asks why a student engages, is barely studied at all.

Key Findings

1. The review included 39 articles, selected from 2,757 records identified in Web of Science (1,142) and Scopus (1,615) when the search ran on 14 April 2026. Eligibility allowed publication from 2021 to 2025, but every included study appeared between 2023 and 2025.
2. The evidence base is concentrated: 25 of the 39 studies were conducted in China and 34 in higher education. Twenty-seven used quantitative methods, nine mixed methods and three qualitative methods.
3. Fourteen of the 20 GenAI-supported learning activities in Chiu's (2024) ChatGPT classification appeared in the corpus. Competence-related activities were the most represented, with seven types identified, against five under autonomy and two under relatedness.
4. The most frequently identified activities were getting feedback on work, searching for information, getting examples, and getting insight into complex problems. Personalized feedback alone appeared in 17 articles.
5. Six activities were absent from every included study: generating review questions, creating challenging problems, generating questions for essays, anticipating ChatGPT's outputs, grading ChatGPT's outputs, and debating with ChatGPT. Each asks the student to judge or challenge the tool.
6. Thirty-five factors were reported in relation to motivation, grouped into GenAI characteristics, learner factors, psychological factors, and environmental factors. GenAI characteristics were the most frequent category. Personalized feedback led at 13 studies, perceived autonomy, competence and relatedness at 12, and teacher support at 9.
7. The review reports no pooled effect size. Experimental and quasi-experimental studies "generally reported changes in motivation or related outcomes", cross-sectional studies reported relationships between variables, and qualitative and mixed-methods studies reported both positive and negative experiences.
8. Measurement was dominated by quantitative questionnaires. Seventeen studies measured basic psychological needs, using instruments that include the Basic Psychological Needs Scale and self-developed SDT-based scales. Engagement was measured in 10 studies, self-efficacy in 5, and enjoyment and grit in 2 each.
9. Motivational regulation was thin. Autonomous motivation appeared in 2 studies, and identified regulation and controlled motivation in 1 each, far fewer than need satisfaction or engagement. The authors treat that gap as the review's central problem, because regulation concerns why a student is motivated rather than how much.

How the review was built

The frame was chosen before the search. SDT holds that motivation follows from the satisfaction of three needs, autonomy, competence and relatedness, so GenAI features such as control over pace, feedback and practice, and interaction with the tool, teachers and peers become candidate supports. The authors state the discipline that follows: the presence of such features "does not necessarily indicate that the corresponding psychological needs are satisfied". A generic engagement framing would have counted the features and stopped there.

The protocol followed PRISMA. Searches in Web of Science and Scopus combined a technical dimension (generative AI, ChatGPT, large language models, AI chatbots), an SDT dimension (autonomy, competence, relatedness) and an outcome dimension (motivation, engagement, intrinsic motivation). Screening removed 793 out-of-window records, 624 non-articles, 385 duplicates and 52 non-English articles, leaving 903 for title and abstract review. A further 845 were excluded, mostly as unrelated to GenAI or SDT, and 11 of the 58 sought full texts were unavailable. Of the 47 assessed in full, 39 were included. Two reviewers screened independently and settled disagreements by discussion.

Coding was deductive for activities, using Chiu's (2024) 20-item classification, and inductive for factors. Thematic analysis followed Braun and Clarke (2022). The study was the unit of coding, and each study counted once per code while still able to carry several codes. That convention is how studies reporting motivation as a lone variable were handled: one construct, one named measure, no item-level detail. Constructs and measures were coded as the included studies reported them, not re-derived.

Quality was appraised with the Mixed Methods Appraisal Tool, chosen because it supplies design-specific criteria for qualitative studies, randomized controlled trials, non-randomized studies, quantitative descriptive studies and mixed-methods studies. Each study was classified by design and scored on the five matching criteria as Yes, No or Cannot tell, with no overall numerical score, following MMAT guidance. Given the heterogeneity, the synthesis was narrative rather than statistical, separating experimental, cross-sectional and qualitative evidence.

The appraisal verdict is qualified but not damning. Most studies met the criteria for appropriate measurement and data analysis. Information was often insufficient to judge sample representativeness, several randomized controlled trials did not report blinding, several mixed-methods studies gave limited information on how the strands were integrated, and one longitudinal study raised attrition. The reviewers conclude that reported frequencies indicate patterns across a literature, not strength of evidence.

SDT forces the question an engagement metric never asks

A Student Engagement measure records that a student used the tool, kept at it, or reported interest. SDT asks a prior question: whether the activity was volitional, and which need it served. That distinction is what lets this review separate fourteen present activities from six absent ones and notice that the absent six are the ones requiring evaluation and challenge. The corpus is built around receiving: feedback, examples, explanations, clarification.

The theory also reframes what counts as a motivational result. Personalized feedback can raise reported competence while leaving autonomy untouched, or it can substitute for the student's own judgment. The review cannot adjudicate that, but its coding scheme makes the ambiguity visible, because each activity is tagged to a need category rather than to an outcome variable. A study reporting higher satisfaction after chatbot feedback reads differently once the activity is coded under competence with the autonomy column empty.

The cost of the frame is that the search itself became SDT-shaped. Terms for autonomy, competence and relatedness were required, so studies that measured motivation with other vocabulary were excluded by construction. The corpus is therefore a picture of SDT-labeled research, not of all research on GenAI and motivation.

The corpus cannot separate autonomy support from removed effort

Autonomy activities in the corpus are mostly information acquisition and knowledge exploration: searching, requesting examples, asking for clarification of difficult content. Eight studies reported GenAI clarifying complex material. Those moves remove effort. Whether they also increase the student's sense of ownership over the work is a different question, and no included study answers it.

The six missing activities are diagnostic. Grading the model's output, anticipating what it will say, debating with it, and generating challenging problems all place the student in the judging position. Their absence means the literature has sampled the uses where the tool does the work, not the uses where the student supervises it. A design that wanted to test autonomy support would start from exactly those activities.

Need frustration is the other blind spot. The review states that included studies paid limited attention to how GenAI may undermine autonomy, competence or relatedness, and recommends studying satisfaction and frustration together. Without that, a reported motivational gain is compatible with a slow transfer of judgment to the tool, and the corpus has no instrument that would tell the two apart.

The measurement critique travels further than any of the findings

The third research question produces the review's most reusable result. Seventeen studies measured the three needs, ten measured engagement, five measured Self-Efficacy, and the rest spread across enjoyment, grit, flow, writing interest and similar constructs. Nearly all of it is questionnaire data, usually collected once, and often with an instrument the study team wrote for the occasion.

The authors note that self-report questionnaires may not capture how motivation develops during actual use, since motivation shifts with task, interaction and support. Their proposed remedy is a mixed design: interviews, learning diaries, observations, and interaction data from the tool itself. They also concede a limit of their own review, that instrument validity and reliability were not examined, nor whether an instrument was SDT-based or fit for its context.

That omission matters when reading the tables. Seventeen studies measuring basic psychological needs is not 17 comparable measurements. The same construct appears under the Basic Psychological Needs Scale, a self-developed 17-item SDT-based scale, a perceived autonomy scale, and an SDT-based interview protocol. Adding those together measures the popularity of a construct, not its effect, and the review is careful to say the frequencies indicate presence rather than importance.

What this means for practice

• Instructors. Design for the six absent activities, not the fourteen present ones. Asking students to grade, anticipate, debate with or challenge model output puts them in the judging position that feedback requests do not, and the review found no study of those uses to lean on.
• Learning designers. Treat personalized Feedback as a competence intervention and check separately what it does to autonomy. The review's competence-heavy distribution is a description of what tools currently do, not evidence that the balance is right.
• Researchers measuring motivation. Report the instrument, its source and its validation, and say whether it is SDT-based. The corpus mixes established scales with purpose-built ones, which makes cross-study comparison impossible even where the construct name matches.
• Program leaders. Read motivational claims from this literature as associations. Most of the evidence is cross-sectional, the quality appraisal found reporting gaps on sampling and blinding, and no study pooled effects.

Limitations

• The evidence base is narrow: 25 of 39 studies were conducted in China and 34 in higher education, so generalization to other systems, sectors and cultures is untested.
• Designs, GenAI applications and motivational measures varied enough that no meta-analysis was possible. Frequencies describe the literature, not the strength of any effect.
• No inter-coder agreement statistic was calculated, and the pre-consensus coding records were not kept, so coding reliability cannot be checked independently.
• The search required SDT vocabulary and accepted the broad term "artificial intelligence", which may have missed differently worded motivation studies and pulled in work not centered on GenAI.
• Six prior reviews of AI and motivation are listed in the paper, and the corpus counted here overlaps theirs, so its findings are not independent evidence.
• Most evidence is cross-sectional or a single post-test, leaving motivational change during a course, and short intervention windows, largely unobserved.

Citation

Pan, Q., Abuhassna, H., & Jiang, C. (2026). Generative AI and student motivation: an SDT-guided systematic review of learning activities, motivational regulation, and measurement. Frontiers in Psychology, 17, 1984018. https://doi.org/10.3389/fpsyg.2026.1984018