Generative AI for Learning Individualization and Student Support in Sustainable Higher Education: A Systematic Review of Creative Agency, Equitable Learning, and Implications for SDG 4 and SDG 10

Synthesis: Aybar-Bellido and colleagues reviewed 61 studies, reported in 63 papers, on how generative AI individualizes learning and supports students in higher education, and then asked what those studies can say about creative agency, equitable learning and the Sustainable Development Goals. The answer is narrower than the framing suggests. What the literature documents is mostly learner-directed use — students choosing when and how to consult a model — rather than system-driven adaptation; support functions cluster around explanation and feedback; and the same studies that report autonomy and idea generation also report overreliance. Methodological appraisal found confounding uncontrolled in every non-randomized design, and the SDG connection is almost entirely the reviewers' interpretation: seven studies framed themselves around SDG 4, one around SDG 10, one named a target, and none measured an official indicator. Sustained learning gains and reduced inequality remain unestablished.

Key Findings

1. Sixty-one studies, unevenly distributed. The search covered Scopus, ERIC, ScienceDirect and the MDPI platform, yielding 3,405 records and, after screening, 61 studies in 63 reports from 37 countries. Publication years were skewed to the present: 33 from 2026, 25 from 2025, two from 2024 and one from 2023. Disciplines were mostly multidisciplinary (22), with STEM and computing (14), language and writing (13), health sciences (7) and teacher education (5) behind them.
2. "Individualization" usually means the student decides. Learner-directed control was coded in 46 of 61 studies (75.4%): students chose timing, focus, prompts or pace. Purpose-built or course-customized systems appeared in ten (16.4%), system-driven adaptive personalization in five (8.2%), and educator-designed differentiation in one (1.6%). The authors are explicit that learner control is not evidence of system adaptation.
3. Support functions cluster around explanation and feedback. Explanation and tutoring appeared in 52 studies (85.2%), writing and task feedback in 40 (65.6%), self-regulatory and emotional support in 24 (39.3%), practice, assessment and simulation in 20 (32.8%), and programming or problem-solving in 17 (27.9%). Functions co-occurred: 53 studies (86.9%) reported at least two.
4. Creative agency results are mixed within the same studies. Critical, reflective or problem-solving thinking was reported in 41 studies, autonomy over creative work in 39 and idea generation in 33 — alongside overreliance and reduced independent thinking in 39 (63.9%), loss of voice or authorship in 8, and reduced originality or homogenized output in 8. In one writing experiment, about 71% of AI-assisted introductions consisted of directly copied model text, and participants reported lower perceived control and ownership.
5. Equity findings are thinner and rarely comparative. Thirty-five studies contributed equity evidence: disability-related accessibility in 11, linguistic support in 10, targeted support for identifiable groups or lower-prior-achievement learners in 9. The reviewers note that these show assistance, not reduced disparities; one dyslexia-focused survey did not record participants' diagnoses, and differentiated materials prepared by prospective teachers were never evaluated with the pupils they were designed for.
6. The evidence base is methodologically soft in ways that matter. Dual appraisal with the Mixed Methods Appraisal Tool reached κ = 0.664 (82.9% agreement). Of 675 applicable judgments, 55.9% were "Yes", 24.9% "Can't tell" and 19.3% "No". Confounding control was rated "No" in all 16 non-randomized appraisals; representativeness was "No" in 18 of 24 quantitative descriptive studies; integrated interpretation was "No" in 15 of 37 mixed-methods studies.
7. The SDG connection is interpretive, not measured. Seven studies framed their work around SDG 4 and one around SDG 10; only one named a specific target (4.5), and no study measured an official SDG indicator. A sensitivity analysis temporarily removing 39 of 61 studies (63.9%) left the themes represented but sharply narrower — system-driven adaptation fell from five studies to one — which the authors read as persistence of themes rather than robustness.

What the review did

The review was prospectively registered (INPLASY202660012) and reported against PRISMA 2020. Four sources produced 3,405 records; after de-duplication and screening, 210 reports went to full-text assessment and 150 were excluded, mostly for lacking relevant educational outcomes (122). Citation searching added three reports, producing the final 61 studies.

Two choices shape what the review can conclude. The first is the decision to code themes rather than pool effects: with this literature, a meta-analysis would have averaged self-report surveys, design studies and a handful of trials into a number with no common meaning. The second is the appraisal, which is reported criterion by criterion instead of as an overall quality score, so a reader can see that the weakest judgments concentrate in exactly the places that matter — confounding, representativeness, and the integration step in mixed-methods work.

Why learner control is not personalization

The headline number, 75.4% learner-directed, is easy to misread as evidence that AI personalizes learning. It is evidence that students use chatbots. A learner who asks for an explanation of a topic they are stuck on, at a time they choose, is receiving flexible support; they are not receiving an environment that models their knowledge and adapts to it. Only five studies documented the latter, and the review's sensitivity analysis leaves one.

That distinction matters for procurement and for research. A platform marketed as adaptive can be a chatbot with a course wrapper, and a study that reports satisfaction with learner-chosen support tells you nothing about whether an adaptive system would teach better. The review's contribution is to hold those categories apart and report how little of the literature occupies the second one.

What the SDG framing adds — and what it does not

Framing educational AI around SDG 4 and SDG 10 invites claims about equity at system scale. The evidence in this review sits at the level of experiences and within-setting differences: accessible materials, translation support, targeted help for named groups. None of the included studies measured an institutional or national indicator, and subgroup comparisons did not establish that inequality narrowed.

The reviewers are careful here, and their caution is the useful part. They distinguish connections they made themselves from what authors claimed, and then from the one study that named a target. A reader who wants to argue that GenAI advances SDG 4 now has a map of how thin that argument's evidentiary base is.

What this means for practice

• Instructors. Design tasks that require students to verify and justify generated content rather than submit it, because overreliance was reported in 39 of 61 studies and direct copying appeared even in studies reporting creative gains.
• Administrators. Ask vendors which of the four individualization categories a product actually implements. "Personalized" covers learner-chosen chatbot use and adaptive systems that model the learner, and this literature finds very few of the second.
• Researchers. Separate assisted performance from independent learning, and compare learner groups over time. The review's most consistent complaint is that perceptions stand in for outcomes.
• Policymakers. Do not accept SDG alignment as an outcome claim. Only one included study named a target and none measured an indicator, so institutional reporting should specify the target and the measure before the claim.

Limitations

The review synthesizes a young literature: 58 of 61 studies were published in 2025 or 2026, which limits how much longitudinal evidence could exist. Coding themes across heterogeneous designs yields frequencies, not effect sizes, and the authors say so; a theme's presence in 39 studies does not mean the effect was large or consistent. The primary exclusion rule recorded one reason per report, so reports failing several criteria may be underrepresented in the flow diagram. The search was limited to four sources and to English-language publication, and citation-searching counts were incompletely recorded. The sensitivity analysis, which retained 22 of 61 studies, narrows dimension coverage enough that the retained set cannot support the same conclusions. Finally, the reviewers report that ChatGPT was used to assist drafting, which does not affect the synthesis but is worth noting for transparency.

Citation

Aybar-Bellido, I., Arones, M., Arévalo-Tuesta, J. A., Adauto-Medina, W., León-Velarde, C., Fernández-Flores, S., Gomez-Zúñiga, F., & Morales-Romero, G. (2026). Generative AI for Learning Individualization and Student Support in Sustainable Higher Education: A Systematic Review of Creative Agency, Equitable Learning, and Implications for SDG 4 and SDG 10. Sustainability, 18(22), 11469. DOI: 10.3390/su182211469.