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

After illustrating the current landscape of the use of generative AI in collaborative learning contexts, the question arises of what the future may hold. As researchers, educators, developers, and policymakers, we must play an active role in shaping this future by utilising educational theories and empirical evidence to design instructional support that benefits learning in small groups. This section presents our thoughts on the current discourse and hopes regarding the development and implementation of generative AI for collaborative learning, but also regarding the necessary research to inform the development of support for collaborative learning.

GenAI to support collaborative learning: stand-alone solution and auxiliary function

While there is already ample research on collaborative learning and adaptive support for collaboration, first hopes on leveraging GenAI to support collaborative learning tend to rely on the use of LLMs as a stand-alone solution for instructional support. Given the rich tradition in adaptive collaboration support, this trend raises an important question: which aspects of the design of LLMs warrant the assumption that LLMs are indeed capable of performing all actions necessary for effective support in collaborative learning?

As illustrated above, to generate adaptive support for groups, generative AI models (and humans) must be capable of performing activities such as those laid out in Molenaar’s detect-diagnose-act framework (Molenaar, 2021). Research in the areas of CSCL, learning analytics, and AI in education have been exploring how to derive indicators from multimodal data streams that represent objective, reliable, and valid operationalisations reflecting relevant characteristics of learners, groups, and collaborative interaction. These efforts have shown that this is no trivial task. Besides conceptual work (Drachsler and Goldhammer, 2020; Wise, Rosé and Järvelä, 2023), there are also empirical approaches that aim at exploring multimodal data streams (e.g. video, voice, computer log files) to detect aspects of the collaboration process (Sottilare et al., 2018; Zhou, Suraworachet and Cukurova, 2024; Järvelä et al., 2021). Notably, the results described in the overview by Schneider et al. (2021) point to persisting challenges to derive valid indicators for different aspects of collaborative learning.

Supposing that the type and goal of the support have already been decided, a GenAI model then needs to determine whether support is required and then generate supporting actions that elicit productive interaction (e.g. by providing feedback, contrasting cases, or prompts). In other words, instead of merely predicting a “next token” (i.e. parts of words or sentences) generative AI systems are assumed to be capable of predicting “the next instructionally adequate actions”. Given criticisms that LLMs represent “stochastic parrots” (Bender et al., 2021) and potentially generate advice that reproduces neuromyths as Richter et al. (2025) illustrate, one may wonder whether it is realistic to expect LLMs to be capable of providing instructionally beneficial support, in the sense of a “next token pedagogy”. In this regard, a thorough discussion among researchers and developers is needed to clarify our expectations of LLMs in terms of their ability to detect relevant characteristics of collaborative learning and groups, leverage this information for diagnoses, and generate adequate supporting actions that address the diagnosed states. This discussion should be accompanied by empirical investigations into the functioning of LLM-based support that compares machine outputs with best practices derived from research on learning and instruction. As we have seen, many researchers specify pedagogical guidelines that the LLMs are expected to follow. However, whether the systems indeed produce the desired outputs in response to events during collaboration or groups’ requests has rarely been reported (see An et al., 2024).

A second approach (and question to discuss) is whether generative AI applications such as LLMs can complement or expand established means of collaboration support. Given the wealth of research on collaboration support, it is surely beneficial to explore how to leverage the capabilities of generative AI in combination with already existing types of collaboration support. The most prominent features of LLMs are their natural language processing capabilities and their potential dialogic nature (if implemented alongside a chat interface). A core value of collaborative learning is the interaction between humans, who interact to support each other and co-construct knowledge. Provided that it does not undermine human-human contact and interactions, generative AI may have its place in collaborative learning. For instance, it may facilitate various aspects of the collaboration (i.e. targets), such as social interaction, socio-cognitive processes, or group metacognition (see above). To this end, the support must elicit beneficial interaction patterns, including sharing information, giving explanations, and monitoring the collaboration (King, 2007; Kreijns, Kirschner and Jochems, 2003; Nokes-Malach et al., 2019).

Here are a few potentially fruitful contributions of generative AI to already established types of collaboration support. First, an LLM may adapt the phrasing of an external collaboration script (scriptlets: Kollar, Wecker and Fischer, 2018) to the competence of the individual group members, for example by using easier terminology or providing suggestions about how the prompt from the script may be performed. If the script is more fine-grained, for instance on the level of individual utterances or sentence-starters (scriptlets: Kollar, Wecker and Fischer, 2018), an LLM may help formulate messages that are tailored to the context of the task and the group’s conversation history. At the same time, there is the potential for groups to interact with and adapt the support during collaboration: if support is delivered by a chatbot, groups may ask clarifying questions, such as how to implement a particular collaboration strategy. Groups may also modify the degree of support, for instance by increasing or decreasing the amount or granularity of support, or by selecting specific aspects of their collaboration for which they wish (or no longer wish) to receive support. In this case, GenAI provides not only adaptive support, but also adaptable support, that is, the group can modify the support based on their needs during collaboration (see Rummel, 2018; Vogel et al., 2022; Wang, Kollar and Stegmann, 2017).

When integrated into conversational agents (Burkhard et al., 2022) or intelligent tutoring systems (Diziol et al., 2010; Walker, Rummel and Koedinger, 2011; Walker, Rummel and Koedinger, 2014), an LLM may render the artificial agent’s text output more human-like, allow for natural language input from the learners, provide prompts that are more context-aware, and take into account the history (i.e. context) of the conversation between the learner and the artificial agent, such as the goals that the group is currently pursuing. Accordingly, learners may engage with the support messages more actively. The same may apply to systems that do not take the role of a facilitator, but of a peer that has to be taught by the group and thus elicits beneficial interaction patterns (regarding teachable agents, see for example Brophy et al.,1999, and Hayashi et al., 2025).

The effects of such applications of GenAI should be subject to empirical inquiry. It is essential to note that outputs from LLMs are likely to suffer from encoded biases and stereotypes (Bender et al., 2021); for examples in learning contexts, see Kotek et al., (2023) or incorrect information that may even include neuromyths about learning and teaching (Richter et al., 2025). Researchers, developers, and teachers must be sensitive to such issues and test a system’s performance before implementing it on a broader scale.

Finally, generative AI, and especially LLMs, may be a valuable addition to the analytics pipeline of adaptive support systems due to their natural language processing capabilities, as illustrated by Wong et al. (2025) and Zheng et al. (2021). In this context, GenAI is not directly responsible for diagnosing the collaboration or deploying pedagogical actions but is part of the analysis of the interaction (e.g. dialogue in the group).

So far, we have discussed various specific ways to leverage GenAI to support collaboration. Next, we highlight some more general aspects to consider when designing support that includes GenAI. Importantly, the design of adaptive collaboration support needs to centre on challenges that groups experience, as opposed to employing GenAI for its own sake. Thus, researchers and developers should design support that targets specific challenges that may arise during collaborative learning. Here, the discussion returns to our assessment of the current state of research on the effects of GenAI to facilitate collaborative learning. While the evidence regarding tangible beneficial effects is currently limited though positive, we can conceptualise collaboration support that might be effective. The mechanisms underlying effective collaboration and the challenges for groups are well documented (overviews of the central theories and empirical evidence are collected in Cress et al., (2021) and Fischer et al. (2018). Kollar et al. (2025) and Trentepohl et al. (2025) illustrated the importance of designing support carefully based on established theories and empirical evidence. Luckin and Cukurova (2019) illustrate one way in which insights and methods from the learning sciences can guide the design of effective instructional support.

We have proposed conceptualising adaptive support as scaffolding. From this perspective, support does not remain available to the groups, but is rather faded out as learners become increasingly competent to perform all aspects of the task on their own (Reiser and Tabak, 2014). The ultimate goal of scaffolding is to achieve learning that develops over time and leads to independent performance, as opposed to increasing the mere performance in the moment (and perhaps only as long as support is available).

More specific guidance in designing support can be found in the framework of CSCL design dimensions presented by Rummel (2018) (see Figure 4.1), to which developers can refer in order to familiarise themselves with design decisions and potential options and subsequently utilise theoretical and empirical work to inform the specific design of the support. One perspective that may be fruitful when reflecting on these more fine-grained design decisions is the concept of cognitive offloading (Risko and Gilbert, 2016). This perspective highlights the question of which (cognitive) activities are delegated (e.g. an object in the environment or an artificial agent) and thus not (yet) performed by human learners. While offloading or delegating certain tasks may lead to an increase in performance (Vaccaro, Almaatouq and Malone, 2024), it should be kept in mind that specific activities represent opportunities for learners, and delegating them may thus have undesired effects (Fiorella and Mayer, 2016) or desired difficulties (Bjork and Bjork, 2020). From this perspective, there are good arguments for on-loading (and supporting) activities such as generating and collaboratively exploring different arguments and explanations or supporting other group members in monitoring and regulating their engagement during collaboration.

Finally, the design, effects, and consequences of supporting learning, especially when leveraging machine learning models, should be considered in the light of ethical, legal, and social implications (ELSI) or aspects (ELSA) (Nakazono, 2023; Zwart, Landeweerd and van Rooij, 2014). Technology is never neutral, and we therefore need to reflect not only upon our use of technology but also its design (ethics by design). Guidance for such reflection might be found in the field of applied ethics (Simis, 2024), and especially from specific ethics such as the ethics of technology (Hansson, 2017), ethics of artificial intelligence (Gunkel, 2024), and ethics of artificial intelligence in education (e.g. Holmes and Porayska-Pomsta, 2023). A more practical approach for ethical reflection is offered by Simis (2024; 2025), who provides suggestions and leading questions. Given the current trends of exploring the use of LLMs in educational settings, we summarise some insights that might be included during ethical reflection: while there are undoubtedly potential benefits of utilising GenAI as part of educational technologies, we must be aware of potential costs (legal, social, or economic) associated with them. Bender (2024) and Bender et al. (2021) describe some of the costs linked to specific LLMs, including the question of which materials companies can legally use to train their models (Brittain, 2025; Creamer, 2025; O’Brien and Ortutaty, 2025) or precarious working conditions of the data workers involved in the training and fine-tuning of models (e.g. Ali et al., 2024; Hao and Seetharaman, 2024). Other costs are biases and stereotypes that are encoded in machine learning models, and the energy necessary to train models and process user queries (see Cornell University, November 3, 2022; Crawford, 2024), or other “hidden costs” (Greenbaum and Gerstein, 2025). With this in mind, we have to discuss which costs we are willing to tolerate in exchange for benefits such as positive effects (of a particular magnitude) on learning outcomes and goals such as the UN Sustainable Development Goals (OECD, 2022; United Nations, 2015), for instance quality education, gender equality, or decent work and economic growth. Here, we have to keep in mind that the costs associated with LLMs can be expected to differ depending on the specifics of the development process (training) as well as our way of implementing them in educational contexts.

Conducting research on the effects of generative AI-based support for collaborativelearning

As we have illustrated, more research on the effects of GenAI support is necessary to better inform the design of collaborative learning settings. Therefore, we sketch out relevant design aspects of empirical studies that are essential for creating robust and sufficiently reliable evidence to inform stakeholders’ decisions about the design of effective collaboration environments.

Our review revealed that greater research attention needs to be placed on roles of support, goals of support, and aspects of the collaboration that instructional support may address to achieve these goals. Thus, one avenue for future research is to explore these more thoroughly. Such research could be guided by the framework of CSCL design dimensions proposed by Rummel (2018). Given the crucial role of collaboration processes for learning, research should also investigate processes that occur in the group, such as which roles the learners in the group take when interacting with each other and with an artificial agent (e.g. Gong et al., 2024; Strijbos and Weinberger, 2010), the quality of the interaction (Strauß et al., 2025) the regulation processes (Hadwin and Oshige, 2011; Järvelä and Hadwin, 2013), or the processes of knowledge co-construction (Popov, van Leeuwen and Buis, 2017). As noted by Cukurova (2025), research should not be limited to cognitive outcomes. Collaborative learning offers opportunities to learn how to build relationships with peers, practice emotional regulation, build self-esteem, or acquire metacognitive capabilities.

However, future research might not only expand on potential targets and goals of collaboration support, or systematically investigate other CSCL design dimensions (Rummel, 2018), but might also investigate the consequences of introducing artificial agents into groups, such as altering the interaction in the group and thus how information is processed by human learners. For example, the review by Vaccaro et al. (2024) highlights that the success of human-machine teams is not guaranteed. Instead, performance benefits of human-AI teams are only more likely if the human alone would outperform an AI alone.

From the perspective of cognitive offloading (Risko and Gilbert, 2016), the introduction of artificial agents may lead human learners to delegate tasks to the artificial agent instead of performing these tasks themselves. This may undermine activities that are essential for learning, especially generative learning activities (for overviews see Brod, 2021; Fiorella, 2023; Fiorella and Mayer, 2016) that pose desired difficulties for the learners (Bjork and Bjork, 2020; Bruin et al., 2023). The same applies to monitoring and controlling the collaboration. For the context of individual learning, Fan et al. (2025) found that some learners delegated metacognitive processes to an artificial agent, a finding they termed “metacognitive laziness”. At the same time, it is important to acknowledge that cognitive offloading is not necessarily always detrimental, and can be part of the fading-out process like Reiser and Tabak (2014) (facilitative) co-regulation of collaboration (Rogat and Adams-Wiggins, 2015).

In pursuing avenues like those described above, we seek to gain insights that will allow us to determine how to design effective collaboration support for different educational contexts.

The basis for designing empirical studies is to test hypotheses on the expected mechanisms underlying how the support affects collaboration processes. Ideally, we can then develop experiments that isolate the effect of this support on outcomes such as knowledge acquisition. Conducting such studies in the context of collaborative learning comes with additional challenges, such as small sample sizes, randomisation, and statistical analyses (for a detailed discussion see Janssen and Kollar (2021). Experiments on the effects of GenAI-based support are especially informative for educational practice if an experiment compares this newly designed support with a “strong” control condition, such as other types of support (e.g. collaboration scripts, conversational agents). A comparison of outcomes from groups receiving adaptive support that utilises GenAI with the outcomes from groups receiving no support or receiving a “business-as-usual” instruction may confound the effects of the support with other factors, such as the novelty factor, and may not be very informative for educational practice (Weidlich et al., 2025). When comparing conditions, it is further vital to check whether the adaptive system indeed provides the intended support (i.e. implementation check). Otherwise, the results provide little insight into how to design and implement adaptive support. This is especially relevant for situations requiring insights about the conditions under which different types of support are effective, enabling us to select between alternatives.

Ideally, there will come a time when ample evidence from rigorous research is available that can be synthesised into overviews such as systematic literature reviews or meta-analyses. This appears to be a challenge pertinent for the current generation of research on the effects of generative AI on learning, as illustrated by Weidlich et al. (2025). Thus, we need more research that yields robust findings.

We can tackle the complexity of developing and evaluating GenAI support for collaborative learning by bringing together expertise from fields such as the learning sciences, artificial intelligence, computer science, and educational practice. Such interdisciplinary collaboration can lead to more comprehensive research designs, more nuanced and robust data analysis, and consequently a deeper understanding of how a specific type of support affords collaborative learning. While such collaborations require more time and effort, we believe that the costs required to generate robust and reliable findings are justifiable given the consequences of premature conclusions (i.e. slow science, Frith, 2020; Weidlich et al., 2025) such as investing time to develop and implement ineffective support, or hampering learning. One framework for conducting interdisciplinary research that aims to exert an impact on practice in authentic learning contexts is design-based research (The Design-Based Research Collective, 2003; Puntambekar, 2018; Topali, Ortega-Arranz and Molenaar, 2026).

Instead of “moving fast and breaking things”, we advocate that the costs for conducting research that produces reliable insights are well spent if harms are mitigated and we are able to have confidence that the support indeed benefits our learners. As researchers, educators, developers, and policymakers, it is our responsibility to shape a future where educational technology is used to afford meaningful collaboration and support effective interaction that serves learning. To this end, technologies need to be designed thoughtfully. As Schleicher (2018) notes, “digital technologies are also creating opportunities that will amplify great teaching, even if great digital technology can never replace poor teaching” (p. 68). Therefore, considering insights from research on collaborative learning is the backbone of the design of support that benefits groups and not only fosters the acquisition of domain-specific knowledge, but also helps them become capable team members.