On this page

No primary data: a critical integrative review synthesizing 77 secondary sources and policy texts, with no sample, instrument or fieldwork of its own.

Synthesis: Komodromos (2026) asks when generative AI can produce durable learning without eroding the economic, social, environmental and institutional conditions learning depends on. From 77 purposively selected database and policy sources, the review defines pedagogical sustainability as learning that endures and transfers under maintainable conditions, then proposes the Sustainable GenAI Pedagogy Framework (SGPF): durable learning value is a non-compensable gate weighed against four condition pillars, while a maturity model operationalizes the institutional pillar across five levels and six dimensions, calibrated to the AI Act, GDPR and EHEA standards. Average effects on achievement, motivation and higher-order thinking are positive but heterogeneous; the language-skill effect is probably inflated by small studies, metacognitive benefit is weakly evidenced, and durability is untested.

Key Findings

  1. Durable learning value gates every other consideration. Learning that endures and transfers beyond the immediate task under economic, social, environmental and institutional conditions that can be maintained over time; an application with no learning benefit cannot be redeemed by being cheap, accessible, green or well governed.
  2. Average effects are real but bounded and design-dependent. Meta-analyses report moderate positive average effects on achievement, motivation and higher-order thinking, but heterogeneity is high, the language-skill effect is probably inflated by small-study and publication effects, metacognitive evidence is weak, and durability after assistance stops is untested.
  3. Scaffolding, not substitution, is decisive. GenAI scaffolds when it prompts questioning, verification and understanding, and substitutes when students delegate the thinking; gains were far larger where teachers supported students during interaction.
  4. Detection tools fail, and their errors land on multilingual writers. 14 detection tools were neither accurate nor reliable, and GPT detectors misclassified more than half of essays by non-native English writers as AI-generated, so detector output should never decide an academic case alone.
  5. Access is not capability, and provider dependence compounds it. Students differ in converting access into constructive use, tiered subscriptions add a pay-to-benefit gradient, and a few firms set terms unilaterally, so institutions need disaggregated outcomes, institution-wide licensing and exit-ready contracts.

What pedagogical sustainability means, and why learning value gates everything

The review folds three common senses — durable learning gains, institutional practice and education systems — into one: a pedagogical practice sustains when it produces learning that endures and transfers beyond the immediate task under economic, social, environmental and institutional conditions that can be maintained without eroding learners' independent capability. Durable learning value means retention, transfer and independent capability; the four conditions are its boundary. Education for sustainable development concerns what learners study, whereas pedagogical sustainability concerns whether a way of teaching keeps producing learning. Learning value cannot be traded away, because it is the benefit against which the four conditions are weighed; against eight existing frameworks, none combines economic and environmental accounting with pedagogy and governance.

What the meta-analytic evidence supports, and where it bends

Several pooled syntheses are rated critically low under AMSTAR-2, so their results are indicative. Chen and Cheung's overall estimate is g = 0.804 (95% CI [0.636, 0.972]) with I² = 93.1% and Egger p = 0.002, falling to g = 0.321 after trim-and-fill; language skills fall from g = 2.331 ([1.617, 3.044], I² = 96.6%) to g = 1.009, and metacognition returns g = 0.078 ([−0.495, 0.652], n.s.), I² = 94.1%. Fan et al. report g = 0.499 ([0.372, 0.627]), Han et al. SMD = 0.45 ([0.43, 0.47], I² = 95%, fail-safe N = 4882), and Gu and Yan g = 0.683 ([0.17, 1.19], I² = 94.4%), where teacher support raised the effect from g = 0.077 to g = 1.426. Effects above 0.20 are among the largest observed on standardized achievement, so the language figure is an upper bound inflated by small-study effects; the metacognition result is a null with uncertainty, and no synthesis tests retention after assistance stops.

Integrity, equity, providers and the conditions that decide durability

Institutions have focused on detecting AI text, but detectors are unreliable, adversarial paraphrasing degrades them further, and false positives fall disproportionately on multilingual students; Annex III of the AI Act lists test-monitoring systems as high-risk, so detector output should never decide an academic case alone. The deeper problem is assessment validity: if a system can produce most of a take-home essay, it may no longer measure reasoning or synthesis. Process-based and authentic designs — viva-style defenses, staged assignments, reflective commentaries, portfolios and documented AI interactions — assess judgment, interpretation and verification rather than outsourced output. Beyond the classroom, capability rather than access decides who benefits, provider dependence is political economy, environmental accounting remains thin, and the EHEA standards make assessment validity a quality-assurance obligation.

What this means for practice

  • Instructors. State the intended outcome, keep an unaided test of it, and make verification visible in the task.
  • Instructional designers. Move from scaffolded prompting to critique and unaided performance; reinvest drafting time in feedback and assessment design.
  • Educational technology developers. Instrument cost per learner, disaggregated literacy and detector errors, provider energy disclosure, and contract terms on data, portability and exit.
  • Administrators. Apply the maturity model before scaling, classify uses under the AI Act, complete a data protection impact assessment where student texts are processed, and treat detector scores as one signal among several.
  • Researchers. Run cross-national panels, longitudinal retention studies, performance-based literacy measures and detector audits by first language.

Limitations

  • Integrative, not systematic: purposive selection and unlogged screening allow replication and updating, but not exact reproduction.
  • Appraisal was single-reviewer, and conclusions inherit secondary syntheses that overlap in primary studies and were mostly rated critically low.
  • English-only, and only 8 of 21 primary studies were European, so European claims rest on regulation and thin contextual evidence.
  • Cost and environmental figures are anticipated, not measured; providers rarely disclose per-query energy data.
  • The SGPF and maturity model are conceptual; the illustrative case shows use, not validity.

Citation

Komodromos, M. (2026). Generative AI and Pedagogical Sustainability in Higher Education: A Critical Integrative Review with a European Application. Sustainability, 18(21), 11239.

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.