A scoping review published in Frontiers in Education has mapped the emerging field of generative artificial intelligence (GenAI) used as a direct pedagogical agent in education for sustainable development (ESD), identifying 25 studies published between 2024 and 2026. The review, conducted by researchers from Universidad de Las Américas in Ecuador and Unidad Educativa Bilingüe Los Arrayanes, found that while GenAI tools are being integrated into sustainability teaching, the evidence for measurable learning outcomes remains limited, and several critical risks specific to the domain remain underexplored.
According to the review, searches in Scopus, Web of Science, and ERIC on 8 July 2026 retrieved 172 records, of which 25 studies (reported in 26 publications) met the eligibility criteria. The corpus is roughly evenly split between Global North and Global South settings, and is concentrated in higher education. The researchers followed the Joanna Briggs Institute methodology for scoping reviews and reported under the PRISMA-ScR extension.
Of the 21 implemented interventions identified, 20 were integrated into instruction rather than running alongside it, the review reports. However, only 11 of those framed the generative activity with briefing, debriefing, or guided reflection. Generative systems were deployed most frequently as dialogic tutors and instructional co-designers, the authors note.
All 25 studies targeted cognitive outcomes, but the review found that measured change in learners was evidenced in only ten of them, and against a comparison condition in only five. The authors argue that this points to a need for more rigorous study designs that can isolate the effect of the generative component from other instructional elements.
The review identifies three converging gaps in the current literature. The first is a gap of framing: it is not yet clear to what extent GenAI has been deployed to develop sustainability competencies rather than simply to widen access to sustainability content. The second is a gap of evidence: no prior synthesis has documented which pedagogical roles generative systems assume in ESD, under which interaction designs, or with what quality of measurement. The third is a critical gap concerning risks distinctive to this technology in this domain, including the flattening of complex socio-ecological systems into fluent single-factor explanations, the delegation of normative judgement to a system that can produce evaluative positions without warrant, and the unexamined environmental cost of the instrument itself.
"ESD is not a domain in which efficiency is straightforwardly desirable," the authors write. "The competencies it seeks to develop are exercised precisely in the difficulty of reasoning under uncertainty and conflicting values, and a system that resolves that difficulty on the learner's behalf may work against the educational aim it is meant to serve."
The review also notes that a separate body of work has begun to examine the material cost of GenAI itself. According to the authors, a recent scoping review of AI for sustainability and sustainable AI in higher education documents consistent evidence that training and maintaining large-scale models consumes considerable energy and generates substantial emissions, while institutional adoption proceeds with limited ethical or environmental guidance. That literature, however, addresses AI as institutional infrastructure rather than as a teaching instrument, and it does not ask whether educators teaching sustainability account for the footprint of the tools through which they teach it.
For university students, the findings carry several implications. If GenAI tools are increasingly used as tutors or co-designers in sustainability courses, students may encounter systems that present fluent but potentially oversimplified explanations of complex issues like climate change or biodiversity loss. The review suggests that without careful instructional design—including briefing, debriefing, and guided reflection—students may not develop the critical competencies that ESD aims to foster. Additionally, the environmental footprint of the AI tools themselves raises questions about whether their use in sustainability education aligns with the values being taught.
The review proposes four evidence-informed design principles to guide future work: explicit operationalisation of sustainability competency frameworks; study designs able to isolate the generative component; delayed measurement of outcomes; and deliberative configurations that keep normative judgement with the learner. It also calls for attention to the environmental footprint of the tools themselves.
The authors recommend that the field consolidate around more rigorous methods, including explicit use of sustainability competency frameworks such as the one consolidated by Brundiers et al. (2021) and refined by Redman and Wiek (2021), or the European GreenComp framework (Bianchi et al., 2022). They also urge researchers to design studies that can isolate the effect of the generative component, to measure outcomes after a delay rather than immediately after an intervention, and to ensure that normative judgement remains with the learner rather than being delegated to the AI system.