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

“We are dwarfs standing on the shoulders of giants,” Bernard de Chartres famously said, characterising the cumulative dynamics of knowledge: new discoveries are primarily elaborations and combinations of past ones. Access to and mastery of existing knowledge is key for researchers to build on that knowledge and make new discoveries. Hypothesis generation is key to the research process and making new discoveries; it is closely tied with the existing knowledge on which it relies, but it also involves distinct mechanisms that will be examined in the next section.

With the mounting number of scientific publications (articles, databanks, images, computer programmes, etc.), it has become increasingly difficult for researchers to keep pace with advances in their own field, despite increasing specialisation within scientific domains. Hence a new challenge for researchers: improving on knowledge with which one lacks familiarity. AI has given rise to tools that can support researchers in these tasks, such as generalist LLMs or specialised models (e.g. Elicit, Consensus, Clarivate, PaperQA2, BioloGPT) (You, 2024). These tools can conduct knowledge management operations like searches in the literature, summaries and literature reviews, which we will examine below. These models work as well for education research although they are not fine-tuned for this domain.

The researcher enters a particular research question in the model (for example, “is virus X responsible for disease Y?”), and the model responds with a list of publications relating to the query, and for each publication, a summary of its results relating to the question. Some models offer a synthetic (consensus) view of the literature, with lists of publications agreeing or disagreeing with the consensus and the corresponding arguments. Some tools can generate a graphic picture of the concerned research landscape, with citations-based relations between publications (who cites whom, who is co-cited with whom, etc.) (Kudiabor, 2024).

Compared to LLMs, specialised models aim to offer superior reliability, as they make use only of scientific publications, avoiding blogs and other sources of lesser scientific reputation. Certain tools can also offer other products beyond the aforementioned search results, like literature graphs (extracting the main concepts or results of a domain and relating them to each other in a knowledge graph). Certain platforms have a “Chat with PDF” function, which allows the user to upload a paper and ask questions about its content (Heidt, 2025).

AI models can produce summaries of publications on request, which gives researchers a rapid overview of a set of papers of interest, gaining time in reading and allowing focus on the most relevant papers. However, the quality of such summaries can be low at times. Peters and Chin-Yee (2025) compared the AI summaries to human summaries that some journals provide for 4 900 examples in medicine and science overall. They found that all AI models tend to over- generalise the results presented in the papers, as they often omit important details that restrict the domain of validity of the results and leave out relevant nuances. They might, for example, just state that a drug is effective for treating a certain condition, without specifying in which dose or for which group of patients. This reflects the difficulty for AI models to fully recognise the importance of “details” that intelligent human readers find significant. The same issue would apply to education where results may be more or less relevant depending on country, socio-economic background, sex, etc.

GenAI models can provide structured summaries of the literature relating to particular questions (Skarlinski et al., 2024). These reviews are useful to researchers for getting a broad view of a question while saving time and making sure that they do not miss the most important relevant publications.

Some models offer “systematic” reviews, which include granular information on the methods and results of each paper in a standardised manner; this is necessary if a researcher wishes to reproduce an experiment or perform a meta-analysis. Some researchers are sceptical about the quality of AI-generated systematic reviews (Pearson, 2024), as AI models tend to skip specific but important information like the precise dose of a drug, as mentioned above. This would be the same for the precise pedagogical context of, say, studies about the impact of project-based learning or lecturing or the use of technology in education.

More generally, literature reviews conducted with AI have certain limitations. First, many models can access only the abstracts of all publications and the full text of open access publications. Access to a significant part of the scientific literature is restricted (although publishers’ tools can, of course, access their own publications), so that many important research findings, and most notably methods, are often skipped in AI-generated reviews.

Second, GenAI literature review tools sometimes struggle to identify the most relevant papers in a field, and to identify topical versus outdated literature, and can first list literature with ideas that used to dominate a domain but have now become outdated.

Such systems could still be used to update human-authored literature reviews rather than generate new ones. Reviews cannot feasibly be updated by humans very frequently and AI could provide for this, even though authoritative reviews may still need human involvement.

Most recent models, like OpenAI Deep Research or Gemini Deep Research, can provide “research reports”, which go beyond a literature review as they provide broader background and context and identify pending questions (Heidt, 2025). The user can enter a query, together with their own data (articles, etc.), and the model returns a full report, including text, figures and corresponding bibliographic references. These models mimic how a person would approach a research question. This is especially interesting when exploring a domain with which the user is not familiar: it helps access general knowledge in a clear language (Jones, 2025).

One specialised model, PaperQA2, writes Wikipedia-style summaries of scientific topics that are cited and significantly more accurate than existing, human-written Wikipedia articles. It can identify contradictions within the scientific literature, a task that is challenging for humans (Skarlinsky et al., 2024).

Certain models generate draft research reports in which they identify gaps in knowledge, getting close to suggesting further possible research topics. The quality of these reports is debated (Jones, 2025): they often include incorrect (or invented) citations, they fail to distinguish authoritative information from simple suppositions, and they do not convey uncertainty accurately.

Hypothesis generation and ideation

Hypothesis generation is a defining activity for a researcher. It consists of generating ideas from the literature or from data which are altogether novel, plausible and testable. Whereas a literature review is about what is known, hypothesis generation is about jumping into the unknown: identifying possible responses to questions that are not answered by the literature, while keeping consistent with established knowledge. This has been, until recently, a preserve of humans. Now AI can also do this. It usually involves three steps: generation of the hypothesis, evaluation/validation (or rejection), and improvement/refinement.

In the case of education research, these techniques could help to explain learning trajectories or some puzzling aspects of education outcomes. AI techniques could combine multiple and remote sources of information to generate original hypotheses. For example, one could imagine AI systems generating hypotheses based on international or national datasets to make hypotheses on the factors explaining the increase or decrease of student outcomes, exploiting the big size of these data sets. But it could also connect these results to other possible sources and point education researchers to possible explanations that are not immediately visible in their data sources (e.g. learning outcomes might increase due to the recent availability or social services that lead to less absenteeism of students or better mental health as their parents get better support).

Most models draw hypotheses from the literature, but some models can also do it directly from the data they are offered to analyse. As compared to humans, AI models have the advantage of a broader knowledge of the literature: not only in the concerned discipline, but possibly in others, accessing more diverse sources (assuming that the model has been trained or can access this knowledge, which may be problematic, as illustrated by the limits of automatic literature reviews). However, AI models are confronted with specific difficulties for automatically drawing hypotheses from the literature: 1) the source texts might not make clear what the problems and corresponding hypotheses are; 2) the link between a problem and hypothesis as stated in the literature might not be straightforward; 3) the novelty or feasibility of a hypothesis can be difficult to evaluate, but they need to be measured and ranked; 4) initially designed hypotheses usually need to be improved so as to strengthen their novelty or feasibility, requiring further operations after the extraction from the literature.

One simple way for a researcher to interact with a model is through brainstorming, with a prompt like “give me ten ideas of mechanisms that could explain how A influences B”. The researcher can also challenge the model, by submitting their own hypothesis and asking the LLM for counterarguments or alternative hypotheses. This simple procedure allows for initial suggestions, but for working out structured and plausible hypotheses, a more articulated approach is required, using specialised models.

One approach is to insert highly structured data into a prompt so as to strictly constrain the model’s response. For instance, a researcher who investigates how microplastics are transported through soil and into groundwater could use a visualisation tool, Research Rabbit. The tool takes a single “seed paper” and generate an interconnected web of research linked by topic, author, methodology or other key features. By inserting its results into an LLM, “it’s possible to query the body of work for hidden links or new ideas” (Heidt, 2025).

Regarding the difficulty for LLMs to link problems and hypotheses from reading a corpus, one solution is to fine-tune an existing LLM so that it can better identify problems and hypotheses in papers. O’Neil et al. (2025) have assembled a database of 5 500 scientific hypotheses (HypoGen), with which they trained an existing LLM. These data are structured in a way that makes it clear what the problem is, what the hypothesis is, and what the chain of reasoning from the problem to the hypothesis is.

Extracting hypotheses from data with AI models is made difficult by the lack of explainability of AI. For instance, the correlation between certain events might be difficult to attribute to particular features of the concerned events. The model will observe that phenomenon A is linked with B, but it could not say whether this is due to feature C or D of these phenomena. The model can see patterns in the data that are not visible to humans, and it might be difficult for the model to translate them in words, as hypotheses that can be understood by humans. Ludwig and Mullainathan (2023) propose a procedure that enables the expression of correlations found by AI in the data in words, so that they can be explicated to humans and tested. For this purpose, they use counterfactuals: they generate synthetic data that exaggerate the correlations found in the initial data, to a point where the concerned pattern becomes visible to humans and can be interpreted.

Inferring hypotheses from the literature might not be enough, or fully satisfactory, as “raw” hypotheses might be insufficiently novel or articulate, too similar to their sources, weakly plausible (not fully coherent with the evidence), or difficult to test experimentally. Thus, a process of refinement of the ideas extracted from the literature is warranted. This is one of the most difficult challenges for AI in science, as it requires both imagination and reasoning capacities: the capacity to improve an idea while keeping its core; to infer logically; to assess the proximity of an idea with the “real world”, etc. A lot of developments are occurring in this domain. The main techniques include: multi-step reasoning (“chain of thought”, requesting the machine to make explicit the steps it is following in its reasoning); reinforcement learning (training the model so that it strengthens its successful features and weakens others); evolutionary computation and multi-agents systems (see below). Models developed since 2024 include one or more of these techniques.

Evolutionary computation is a technique inspired by the mutations and natural selection in Darwinian evolution. It begins with a review of the literature, from which it extracts an initial list of hypotheses. It applies small, random changes to an algorithm and selects the ones that improve the model’s efficiency. To do so, the model conducts its own “experiments” by running the algorithms and measuring how well they perform. Afterward, the model produces and evaluates a paper. After “augmenting the literature” this way, the algorithm can then start the cycle again, now building on its own results (Castelvecchi, 2024).

Agentic AI is being applied also to science. An agent is an autonomous system that can pilot various tools towards a given objective. Multi-agent systems are comprised of several agents with specific objectives and specialised skills; each one pilots an AI model (e.g. an LLM) and interacts closely with others under the supervision of a “lead agent”, who acts like the conductor of an orchestra. Some models also integrate reasoning capacities like the aforementioned “chain of thought”. A multi-agent model aims to function like a group of researchers. Some researchers are specialised in a particular discipline; some researchers play a particular role, of making proposals, or challenging others’ proposals or combining them; at each step they are assigned specific tasks by the lead, and they work with their respective tools to implement them; they meet and confront their respective findings, with open discussions whose conclusions are included in a report shared with human researchers. The whole process starts with a prompt that includes a description of the problem and contextual information, which is submitted to the model. The lead and the agents then design a research plan, possibly with a sequence of sub-questions, a set of parallel tasks, a list of skills required from agents, etc. Then an iterative process can take place, in which each agent accomplishes the tasks it has been charged with, and reports to the lead and to other agents; the lead synthesises the findings at each stage and monitors the advancement of the whole process. At the end, the model can draft a research report (Biever, 2025).

In education, multi-agent models with GenAI agents are for example used to develop assessment items. They may hold promise for educational research, which is often multi- or inter-disciplinary or address broad-based socio-technical issues (such as the adoption and use of AI in education). Education researchers may use such models to ensure different types of expertise, information sources and constraints are brought together, for example to generate new ideas on children’s school rhythm, which involves expertise in children’s biological and psychological development and needs, learning science and pedagogy, parental work schedules, etc. (Figure 13.4). One can also imagine some fruitful uses to generate ideas or improve usual hypotheses for education policy research, with some of the AI agents playing the role of different education stakeholders and providing ideas on addressing some education policy issues such as the provision of equal opportunities. The multi-agent models could then make new suggestions of educational interventions or of policies in this simulated environment.

Box 13.2 presents some examples of (specialised) scientific multi-GenAI agent models.

When provided with a research goal in natural language, the system searches and analyses relevant literature, summarising and integrating existing work to generate new research hypotheses and experimental protocols for later validation. The co-scientist supports its recommendations by citing key sources and explaining the logic behind its suggestions (Gottweis, 2025).

Given a background problem description, models first dynamically retrieve inspirations from past literature in the form of related problems and their solutions, along with contexts from a scientific knowledge graph. These retrieved inspirations serve to ground the generated ideas in existing literature. Generated ideas at this stage are not necessarily new. Models have the capability to iteratively boost the novelty of their generated ideas. Given an idea generated by the LLM model compares the idea with existing research in the literature; if it finds strongly overlapping research, the model is tasked with updating its idea to be more novel relative to prior work (much like a good researcher would do). This model also introduces an in-context contrastive model which encourages novelty with respect to background context (Wang et al., 2024).

Swanson et al. (2025) developed the Virtual Lab, a multi-agent model that tasks different AI agents to accomplish different tasks under the supervision of human researchers to perform inter-disciplinary science and investigate broad, complex research questions. First, the human researcher defines two core AI agents: the Principal Investigator (PI) and the Scientific Critic. Based on a short project description, the PI creates additional AI scientist agents, assigning them specific expertise, goals and roles. The human-AI collaboration then happens through two types of workflow: team meetings and “bilateral” meetings. The authors describe the team meeting as follows: first, the human researcher sets an agenda for the meeting, then the PI opens with initial thoughts and guiding questions, after which each scientist agent contributes in turn. The Scientific Critic reviews their input, and the PI synthesises the discussion, posing follow-up questions. After several discussion rounds, the PI generates a final summary and conclusion for the human researcher. In the case of “bilateral” meetings, a designated scientist agent responds to the agenda set by the human researchers, receives critiques from the Scientific Critic, and iteratively improves its answer. After multiple rounds, the AI scientist delivers a refined final response.

AlphaEvolve is based on the (Google) LLM. Each task begins with the user providing a question, evaluation criteria, and an initial solution. The LLM then generates hundreds or even thousands of possible variations. An evaluator algorithm scores these alternatives against the defined metrics, and the best-performing ones guide the LLM to propose new ideas. Through this iterative process, the system gradually develops a population of increasingly effective algorithms (Gibney, 2025).

How effective are GenAI models in generating, refining and evaluating scientific hypotheses? The evidence is still scarce, as testing such models is complex and costly. The most effective tests are implemented by performing actual research and examining the model’s achievement. Diagnoses are diverse due to the diversity of models, the diversity of research questions and the diversity of testing methods.

On broad research questions, the models seem able to provide useful suggestions that point to potentially fruitful research directions. This is due to their very good access to the literature and their ability to process it in a highly structured way. According to Anthropic (2025), their “internal evaluations show that multi-agent research systems excel especially for breadth-first queries that involve pursuing multiple independent directions simultaneously.”

When it comes to more specific research questions, the evidence is mixed. There have been certain impressive achievements, with models able to identify and describe hypotheses that were then successfully tested by researchers (see Box 13.3), but also less successful cases. Successful cases come with significant involvement of humans in the process.

Certain studies also point to the tendency of some models to propose solutions which are plausible but not really novel, including some that have been already explored and abandoned in the past. Wang et al. (2024) conducted extensive evaluation experiments using human annotators with domain expertise to assess the proposals of a multi-agent model called Scimon (Box 13.2). They found that “ideas still fall far behind scientific papers in terms of novelty, depth and utility – raising fundamental challenges toward building models that generate scientific ideas.”

Regarding the AI Co-Scientist (Box 13.2), specialised in research on machine learning: “the authors admit that the papers that AI Scientist has produced contained only incremental developments. Some other people were scathing in their comments on social media. ‘As an editor of a journal, I would likely desk-reject them. As a reviewer, I would reject them,’ said one commenter on the online forum Hacker News.” (Castelvecchi, 2024)

In mathematics, there is also contradictory evidence. Some tests point to the fact that certain of the claimed achievements of some models in solving Olympiad silver or gold medal level problems were due to “data leakage” (Olympiad is a global competition in maths): solutions had been published before and were accessed by the models (Petrov et al., 2025). On the other hand, very rigorous testing by professional mathematicians showed that o4.mini, an OpenAI reasoning model, can solve most of the PhD-level problems they submitted, demonstrating extremely powerful reasoning capacities. However, there is consensus that current models are not up to the level of mathematical research, although they are getting closer (Chiou, 2025).

Two further caveats need to be mentioned. First, not all negative results of testing are necessarily published, and limitations of the models could be underestimated; second, it is difficult to estimate the quantity of human involvement in the models’ work, but it could be sometimes important, and the role of the models could then be over-estimated. However, these models are still in a very early stage of development and much progress is to be expected in the near future.

The AI Co-scientist was tested on a question that took years to resolve: how cf-PICIs (a defence mechanism against phages) spread across bacterial species. It generated and tested several hypotheses. Its top-ranked hypothesis matched the experimentally confirmed mechanism. Other hypotheses opened new research avenues, showing AI’s potential as a creative engine for scientific discovery. Generated hypotheses were novel, logically coherent and experimentally testable, highlighting the system’s ability to accelerate discovery and build transdisciplinary bridges (Penadés et al., 2025).

The Virtual Lab tackled the design of nanobodies binding to new SARS-CoV-2 variants. It combined several models – ESM, AlphaFold-Multimer, and Rosetta25 – to mutate nanobodies targeting the spike protein. Of 92 designed nanobodies, over 90% were expressed and soluble, with two showing unique binding to recent variants JN.1 and KP.3. This demonstrates the potential of AI-human collaboration to deliver complex, validated scientific results (Swanson et al., 2025).

AstroAgents is a multi-agent system with eight AI agents generating hypotheses in astrobiology. Using LLMs (Claude Sonnet 3.5, Gemini 2.0 Flash), it analysed meteorite and soil mass-spectrometry data, producing 101 hypotheses from Gemini and 48 from Claude. Gemini’s ideas were more novel but error-prone; Claude’s clearer but less original. Overall, the system produced plausible, pattern-spotting insights beyond human capability, useful for analysing samples such as those planned for return from Mars (Biever, 2025).