On this page

Synthesis: Bubnys, Butvilas, Colomer and Cañabate read 49 scholarly and policy sources and argue that "sustainable" educational technology has been reduced to a claim about access and efficiency, when sustainability should also mean learning that lasts, inclusion, rights, institutional feasibility and proportionate use of resources. From that synthesis they propose SAREL (Sustainable AI-Supported Reflective Education and Learning), a five-stage design-and-audit loop — Specify, Allocate, Require, Evaluate, Loop — for designing GenAI-supported reflective work in higher education. The framework's contribution is integrative: it connects existing concerns about keeping reflection as learner work, accessibility and agency, governance and workload, and resource intensity into one task-level sequence. The authors are explicit that SAREL, its rubric and its "sustainability-adjusted learning value" construct are provisional and have not been empirically validated, so the review is a design argument rather than a test of one.

Key Findings

  1. Adoption evidence is not learning evidence. Technology-acceptance research measures perceived usefulness and ease of use, which explain uptake but do not establish durable, transferable learning — a tool can be easy, popular and still add little.
  2. Five domains must be read together. The review organizes the literature into pedagogical, agency-and-inclusion, ethics-and-rights, institutional, and environmental/economic domains, and argues each needs its own observable evidence.
  3. Reflection has to remain the learner's work. AI can act as a prompt, mirror, challenger or feedback partner, but if it supplies interpretations the learner has not attempted, it substitutes for the generative effort reflection is meant to build.
  4. AI must beat the non-AI alternative, not just the absence of AI. Structured LLM-supported reflection improved confidence and scores in two early randomized field experiments (N = 145 and N = 112), yet comparable gains came from well-designed non-AI questionnaires; a mathematics study found ChatGPT-generated and human-authored hints produced no significant difference in learning.
  5. Sustainability is a comparative design question. SAREL treats accessibility, rights and learner agency as prerequisites — not scores a strong performance elsewhere can offset — and asks whether a less resource-intensive arrangement would deliver equivalent learning.

The evaluation imbalance

Across the selected literature, justification for adoption clusters on access, flexibility, personalization, efficiency and user acceptance. The authors do not dismiss those constructs; they argue they answer a different question. Reviews of online and AI-supported learning, they note, find outcomes turn on pedagogy, course design, teacher presence and context rather than on the digital medium alone, and find uneven attention to educational theory, educator roles, methodological rigor and longitudinal evidence. The result is an evaluation imbalance: a technology can be convenient and widely used while producing limited transfer, shifting monitoring work onto instructors, or displacing the effort that would have produced learning. The unit of appraisal, they argue, is the whole socio-technical arrangement — task, learners, instructors, institution, platform, data practices and infrastructure — not the tool in isolation.

Keeping reflection as learner work

Reflective learning is a recursive process in which learners revisit experience, surface assumptions, interpret evidence, connect theory to practice, regulate emotion and decide how to act differently. Its value lies in those cognitive and metacognitive operations, not in the fluency of the finished narrative. AI can support the cycle when it asks a learner to justify a conclusion, generates a counterexample, or compares an account against a rubric. The pedagogical risk is cognitive substitution: polished reflective text generated before the learner has articulated a position. The authors draw on cognitive-load and multimedia-learning scholarship to separate useful reductions in extraneous processing from the removal of generative activity. For tasks meant to build independent reasoning they propose three design moves — have learners commit an initial account before consulting AI, treat AI output as an object for critique rather than an answer, and document the visible transformation between the first and final position — which become SAREL's Require stage.

Agency, accessibility and equivalence

Digital inequality, the review notes, is not just device ownership; it includes connectivity, skills, confidence, language, disability access, privacy, household conditions and the ability to convert access into outcomes. GenAI can lower some barriers (translation, simplification, multimodal explanation) and widen others (paid tiers, bandwidth, recent hardware, dominant-language proficiency, prompt skill). Universal Design for Learning supplies the design stance: multiple means of engagement, representation and expression, with meaningful alternatives — text or audio reflection, AI-assisted or non-AI pathways, synchronous human support, transparent criteria that do not reward technical fluency unrelated to the objective. The principle is equivalence without sameness: different media or levels of automation may be used, but they must offer comparable chances to meet the same outcome, and inclusion must be designed in rather than treated as an after-the-fact accommodation. This is the Allocate stage, and it treats applicable rights, learner choice and essential accessibility as non-compensatory prerequisites.

Rights, governance and the workload nobody counts

AI-supported education rests on data, models and institutional arrangements learners cannot see. The review gathers familiar learning-analytics concerns — consent, privacy, classification, surveillance, persistence of student profiles — alongside GenAI-specific ones: uncertain training-data provenance, hallucination, embedded bias, third-party retention of prompts, intellectual-property ambiguity and models that change mid-course. Ethical sustainability, it argues, needs task-level disclosure (which system is permitted, what data must not be entered, what needs citation, how errors are checked, who is accountable) and human review for consequential judgments. Institutional viability depends on workload: automation can cut repetitive feedback while creating new work in prompt design, output checking, exception handling, integrity redesign, training and procurement. A pilot that looks efficient because instructor verification time is excluded, the authors warn, has shifted risk rather than reduced cost. Governance is proposed at three levels — course, program and institution — so individual teachers are not left to solve legal and technical questions alone. SAREL does not invent these principles; it connects existing responsible-AI and risk-management guidance (including the NIST AI Risk Management Framework and updated European Commission guidance for educators) to the reflective task.

Resource proportionality

The environmental argument is subtler than "green IT". Because task-level energy and emissions data are usually unavailable to educators, SAREL treats resource proportionality as an explicit design and comparison question rather than a claimed saving. Digital sufficiency — asking whether a task, its scale or its frequency is necessary, and whether a lighter arrangement would meet the purpose — is offered as a more practical principle than abstract efficiency: a static rubric over repeated queries, a smaller or institutionally hosted model where appropriate, no compulsory high-bandwidth media, longer device lifetimes, less redundant generation. Economic fragility (subscription costs, vendor lock-in, unpredictable pricing) belongs in the same domain, with a low-technology fallback and an exit strategy as design requirements.

The SAREL framework

The five stages are an author proposal, not an empirically derived model:

  • Specify — name the authentic reflective capability and a comparator before choosing a tool.
  • Allocate — specify learner control, accessible formats and an equivalent non-AI route.
  • Require — require reflection and verification, with observable learner interpretation and revision.
  • Evaluate — assess learning, equity, rights, staff workload and resource intensity together.
  • Loop — feed the evidence back into redesign.

A companion table maps five evidence domains — pedagogical, agency and inclusion, ethics and rights, institutional, and environmental/economic — to illustrative indicators: initial-to-final reasoning change and delayed retention for learning; participation gaps and learner control for inclusion; task-level disclosure and appeal routes for rights; total workload and fallback plans for institutional viability; query/model intensity and lower-tech comparators for resource use. The authors frame the goal as capability after assistance: a sustainable intervention should leave learners better able to judge, reflect and act when the tool is unavailable.

What this means for practice

  • Instructors. Before adding an AI step to a reflective task, name the reflective capability you want and a non-AI alternative it must beat; have students commit an initial account and document the change.
  • Instructional designers. Build the equivalent non-AI route and accessible formats in from the start, and treat rights, learner choice and essential accessibility as prerequisites rather than scores that strong performance elsewhere can offset.
  • Faculty developers. Prepare staff to check what AI-supported reflection actually produces — retention, transfer, evidence-checking and calibrated reliance — not just whether the task was completed.
  • Administrators. Count the hidden workload (prompt design, output checking, exception handling, procurement, training) and the resource intensity alongside verified learning value before scaling.

Limitations

The paper is candid about its limits. It is a concept-driven integrative review, not an exhaustive or reproducible systematic review; the purposive selection of 49 sources may carry selection bias, and the screening counts, database queries and inter-coder reliability are not documented well enough for replication. The stated source-type proportions are retrospective classifications, not quality appraisals, and the synthesis is not a weighted meta-analysis of outcomes. Neither SAREL nor its provisional rubric nor the sustainability-adjusted learning value construct has been validated. The authors propose a phased validation program — expert content assessment, feasibility pilots with inter-rater agreement, comparative longitudinal and mixed-method tests, and independent replication — and list candidate measures from immediate and delayed performance to calibrated AI reliance, participation equity, workload and estimated resource intensity. Until then, the framework is a way to frame the right questions, not evidence that answering them this way improves learning.

Citation

Bubnys, R., Butvilas, T., Colomer, J., & Cañabate, D. (2026). Sustainable AI-supported reflective learning in higher education: An integrative review and the SAREL design framework. Sustainability, 18(22), 11812. https://doi.org/10.3390/su182211812

Connected Concepts

Embed this page

Copy the code below to embed this page in a learning management system or other website. The embedded version shows only the page content — it hides the site header, navigation, and footer.