OECD Digital Education Outlook 2026: Exploring Effective Uses of Generative AI in Education

Introduction

This chapter examines the emerging role of generative AI (GenAI), some of the techniques on which it builds and its AI predecessors, in back-end functions of higher education, including course articulation, student transfer, advising, admissions, and content infrastructure. Unlike instructional uses of AI, which often focus on the learner as the end user, the systems discussed here are typically administrator- or staff-facing or are embedded into educational platforms installed at an institutional level to produce insights, reduce task complexity, and support academic pathway navigation (Kizilcec et al., 2023). In most cases, the AI models behind the tools utilise data collected at the macro- or meso-level as opposed to the micro-level (e.g. learning process data; Fischer et al., 2020).

Drawing on recent research, case studies, and early-stage prototypes, this chapter identifies how AI can:

• Support credit mobility and transfer prediction across institutional boundaries • Support academic advising, such as with personalised course and major recommendations and curricular analytics • Diagnose novel opportunities to enhance admissions and resource allocation, and • Structure the classification, tagging, and reuse of learning content and curricular components

While not all of these tools involve generative AI as means to create content or directly interface with end- users, many depend on machine learning, natural language processing, and representation learning (e.g. embeddings), technology at the heart of generative AI, to support institutional decision-making. The chapter foregrounds macro- level, institutional infrastructure as the critical site of innovation for unlocking more evidence-based, personalised, data-informed, and ultimately student-serving higher education ecosystems. While the chapter focuses on research carried out at the higher education level, many of the covered possibilities are relevant to the secondary school sector, at the system rather than institutional level, as well as to support lifelong learning.

Emerging opportunities

Easing learning mobility through AI equivalency models

As students traverse academic pathways, their ability to have learning acknowledged when moving between segments or systems of education can be the difference in their ultimate academic success. In the United States, when students move from a 2-year community college to a 4-year university, for example, agreements called course articulations dictate how much credit will come with them and which requirements it will satisfy. Similarly, prior learning gained in industry from a professional certificate that is then attempted to be counted as equivalent to institutional course credit is referred to as Credit for Prior Learning (CPL) (Lakin, Seymour and Crandall, 2015). In other countries the same issue can arise when individuals want to change study paths, to transition from a 2-year study programme to a bachelor’s degree, change higher education institution outside traditional study paths, etc. This can also happen for the international recognition of foreign degrees in the frame of international student mobility – or just professional mobility. Demonstrating mastery of a skill from one taxonomy and then seeking acknowledgement of mastery in a similar skill from another taxonomy requires mapping, or cross-walk between taxonomies (Li et al., 2021).

These variations on credit and learning acknowledgement scenarios are critical to student success in higher education; however, they have historically been constructed and maintained by hand, often with missing or inequitably distributed pathways for mobility that favour credit from institutions with higher socioeconomic standing (Pizarro Milian and Aurini, 2025; Goulart and Pardos, 2025). Generative AI and the natural language processing technology behind it could and is increasingly being used to address these deficits in ways that have the potential for equitable scaling.

One promising direction to better map different types and levels of educational programmes is to identify closeness across courses with AI techniques. This involves representing course content as AI vector embeddings (see Box 11.1), enabling semantic similarity comparisons across thousands of courses. This representation can be informed both by natural language signals, such as a title and course catalogue description, but also by historic enrolment data, that is, the actual choices courses that individual students made within their higher education programmes. Using the latter, Pardos and Nam (2020) visualised the semantic topology of courses offered at a large public university (Figure 11.1) and queried the underlying course vector representation to reveal differences between courses. For example, when asked what the difference was between the Econometrics and Advanced Econometrics courses, the model correctly responded with “Linear Algebra.” It perhaps would not be surprising for contemporary LLMs to be able to answer this question, given their access to troves of data; however, the model from this example used only course enrolment histories, showing the effectiveness of these models with a better defined amount of data. This effectiveness of course vectors was later demonstrated in successfully predicting student course workload perceptions (Borchers and Pardos, 2023) and also performed strongly on prerequisite prediction (Recall@10 = 0.70) and average-enrolment prediction (RMSE = 42.48) as described in Jiang and Pardos (2020).

Historic enrolment data within institutions can also be leveraged by AI to learn and provide course recommendation pathways using the same type of neural networks which power generative AI (Shao, Guo and Pardos, 2021). Much like a Generative model could complete your sentence, this similar model applied to course enrolments can complete a student’s course sequence to include necessary requirements and nurturing budding personal interests to satisfy electives.

The projections were produced by reducing course vectors to 2-D using t-SNE. The space may suggest to a Dean or other administrator, where a department may have a concentration of topical strength and in what areas “neighbouring” departments may be collaborated with to fill in gaps of a major or work together on a major. If another institution’s course vectors were to be overlaid on to this one, it can suggest where the institutions complement one another, where they are aligned, and where expected alignment could be improved.

An embedding can be thought of as the organisational structure that AI uses to map out what it knows about a domain. Neural networks, the type of AI behind ChatGPT and self-driving cars, ”think” in terms of vectors in this embedding. They take an input (e.g. an image) and convert it into a vector that can then be manipulated through a series of matrix operations (e.g. an image being scaled or rotated as a result of multiplications). Modern AI is usually trained to predict the next word in a sentence, or more generally the next item in a sequence, or to fill in the blanks in past sequences of data. For AI to do this efficiently, it has to pick up on patterns and similarities that allow it to generalise what it has observed in the past (i.e. training data) to new scenarios. It does this by organising the embedding “space” in ways that recognise these patterns, for example by placing synonymous words close together in the space or by giving pairs of words with the same relationship to one another (e.g. Paris -> France, Seoul -> South Korea) similar relative distances to one another in the vector space (e.g. a capitol’s vector in the space might be approximately two units down and one to the left of its country vector in the space). This spatial representation is how neural networks encode what they have learned (i.e. generalised) from the data. When they overgeneralise, this can manifest in hallucinations (e.g. when the expected capitol vector is not at the expected offsets from the country and the AI subsequently gives the wrong answer).

A vector space is commonly hundreds to thousands of dimensions in length. As of this writing, OpenAI offers 1 536 length vectors for its small model and 3 072 for its large. Utilising vectors (i.e. generating and downloading them) in higher education workflows, instead of querying an LLM, is a way of utilising the power of modern AI without having to rely on real-time querying and incurring ongoing API costs. This is a practical approach when a natural language interaction is not the goal. Free vector embeddings are available from the extended research community (e.g. http://www.sbert.net).

Beyond single institution course planning and recommendation, AI can also support the development and maintenance of cross-institutional course equivalency models - an important enabler for student transfer between degree programmes in higher education and a facilitator for lifelong learning and the recognition of prior learning. For example, in the United States, starting at local, more affordable community college and then transferring to a bachelor’s programme has been the greatest source of upward social and economic mobility (Johnson, 2020). Similarly, in the European Union, the European Credit Transfer and Accumulation System (ECTS) is designed to promote student mobility by standardising how learning achievements are measured and recognised across institutions (Grosges and Barchiesi, 2007). In practice, however, the reality of exchange programmes and institutional transfers often involves negotiations for the accreditation of specific courses toward degree requirements. In both cases, course equivalency agreements between institutions are required to allow for transfer to work as intended (Ignash and Townsend, 2000). Across many higher education systems, articulation and credit transfer remain time-consuming, manual processes. Faculty or articulation officers typically review syllabi and catalogue information to determine equivalency across institutions. AI, particularly natural language models and course embeddings, has begun to offer data-driven alternatives and support structures.

The same embedding models that underpin generative AI-used to represent meaning – rich relationships between words, images, and multimodal content – can also represent relationships between courses at different institutions. Pardos, Chau and Zhao (2019) demonstrate that machine translation techniques, built on these embeddings, can “translate” between the course vector spaces of different colleges. These vectors, learned from students’ historical course or programme-level enrolment patterns and course catalogue descriptions, capture latent curricular structures, enabling the prediction of equivalences and surfacing gaps in transfer agreements. These gaps correspond to equivalences that could have been offered given course contents and pathways but that had not been identified or considered yet. In their proof-of-concept, the approach successfully matched courses between a two-year and a four-year US institution and validated 65 pre-established articulations. This methodology is being piloted with 59 US higher education institutions and four systems of higher education to explore its feasibility and utility in practice.

Methodologically, the educational data mining community has explored additional neural course representations outside the transfer context. Khan and Polyzou (2024) evaluated session-based methods such as CourseBEACON and CourseDREAM (neural architectures that recommend well-suited course bundles based on enrolment sessions) and showed improved performance of these methods over traditional factorisation or association models. These session-based models recommend full next-semester sets of courses by modelling (1) which courses pair well together and (2) semester-to-semester orderings using RNN/LSTM encoders (CourseBEACON uses an explicit co-occurrence matrix; CourseDREAM learns latent basket vectors). They improve accuracy over popularity and sequential baselines (CourseDREAM achieves the best Recall@k on test of about 0.30). Similarly, Kim et al. (2025) demonstrate that deep embeddings of course descriptions, coupled with traditional classifiers, can automate equivalency judgments with near-perfect performance (as measured by F1 scores). Both works demonstrate how representations of courses can be used for various institutional tasks guiding student pathways and transfer mobility.

While AI-assistive equivalency models can dramatically speed up articulation, adoption depends on trust-especially among domain experts that ultimately hold the keys to unlock credit approvals. Xu et al. (2023) studied algorithm aversion in higher education administrators tasked with course credit decisions. Using a 2×2 experiment with an AI-based matching platform. One factor was whether low-confidence or outlier AI recommendations were inserted into the results or not. The other factor was whether the interface prompted users to flag inappropriate AI recommendations or not. They found that not including outlier recommendations improved acceptance and productivity; however, asking users to flag recommendations reduced administrators’ acceptance of the suggestions unless outlier recommendations were turned on. While the literature suggested user flagging as a means to increase adoption, it may have led to a negative mindset in this case, unless users were given clear results worth flagging. The takeaway is perhaps that without the right implementation, even accurate AI recommendations may risk being undervalued or outright rejected in a socio-technical system.

These findings show that embedding-based methods, central to modern generative AI, are not only useful for producing language or images, but that they can also map complex academic structures across institutions. When combined with careful human-AI collaboration design, they can accelerate equivalency mapping, reveal hidden curricular alignments, and reduce administrative burden while keeping human experts in control (Kwak, Adelkar and Pardos, 2026).

While many AI applications focus on student-facing outcomes, others operate behind the scenes to power the discoverability, classification, and reuse of educational content. A growing area of impact is the back-end organisation of large-scale resource libraries through tagging, aligning, and curating learning materials to match institutional or state-wide taxonomies or support transition and alignment between them. The significance of using AI for annotating and grouping educational content is that content standards are often changing over time (e.g. the U.S. Common Core or the Finnish National Core Curriculum for Basic Education). Recategorising content and aligning it with new standards is expensive, but could be made substantially more cost-effective and efficient using AI. For example, this can work for Open Educational Resources (OER). While many countries and international organisations like the OECD and UNESCO have supported the development of Open Educational Resources, the COVID-19 pandemic highlighted that identifying those aligned with country or jurisdictions’ curricula was not trivial, because of a lack of domestic or international taxonomy (e.g. Vincent-Lancrin, Cobo Romaní and Reimers, 2022). The strength of generative AI techniques at classifying and mapping text (with embeddings) may help to solve or at least mitigate that problem.

Another example is the conversion of a mastery profile from the proprietary skill taxonomy of an intelligent tutoring system into the US state common core standards, a set of agreed educational standards in English language arts and mathematics across a large number of States in the United States. Intelligent tutoring systems supporting the acquisition of content and procedural knowledge in language arts, maths, science, etc., have their own “knowledge maps” that are often not specific to any particular curriculum or standards. Allowing these translations is essential to help teachers and educators using particular systems to align them with local curricula. These infrastructure-oriented uses of AI may be invisible to learners, but they are foundational for enabling efficient resource retrieval, supporting instructional planning, and ensuring alignment with evolving curricular goals (Figure 11.2).

Recent advances apply techniques central to generative AI, particularly embedding models, to create rich vector representations of learning resources that capture semantic relationships between content items. These embeddings support clustering, that is, the grouping of content according to common characteristics, similarity search, and cross-walking between taxonomies, while classification algorithms, often fine-tuned on top of embeddings, map resources to categories in both established frameworks and newly defined skill taxonomies. For example, when a new mathematical curriculum or taxonomy, like the US Common Core Standards, are introduced, these methods can aid in re-mapping the estimated millions of existing open educational resources to the new taxonomy. Such methods have been deployed to support initiatives like common course numbering, enhance tutoring systems’ ability to link resources to specific knowledge components, and keep course catalogues aligned with rapidly changing programme requirements.

Research suggests that these AI-assisted systems can approach or even match non-expert human tagging performance with relatively small, labelled datasets, and in some cases rival expert performance at scale. For example, Li et al. (2024) found that their approach combining embedding and classification could achieve non- expert accuracy with as few as 100 labelled examples, and near-expert accuracy with 5 000. Importantly, these models incorporated multimodal features from text, images, and videos-mirroring the multi-input capabilities of contemporary generative AI systems-and were publicly released for use with both the US Common Core and novel taxonomies.

Compared to humans working alone, AI suggestions reduced tagging time by roughly 50% (p ≪ 0.01) but led Ren et al. (2024) extended this line of work to study human-AI collaboration in taxonomy alignment.

to a modest decline in recall, that is, the identification of all relevant resources of a specific category (- 7.7%, p = 0.267), and substantial decline in accuracy, that is, the overall correctness of tagging suggestions (-35%, p = 0.117). Notably, the AI-alone condition performed worst, while the human-alone condition performed best for accuracy-placing the collaborative condition in between. These findings highlight a trade-offbetween efficiency, here speed of tagging, and precision, and suggest that while AI can accelerate large-scale taxonomy updates, quality assurance remains essential – as of now, performed by humans.

As educational taxonomies continue to evolve, whether through new competency frameworks, new curricula or standards, or institutional redesigns, embedding-powered alignment tools offer a scalable way to re-tag resources, identify content gaps, and maintain interoperability across systems. In doing so, they extend the same representational methods powering generative AI into the critical, though less visible, infrastructure that underpins curriculum management.