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

As we have illustrated, GenAI based systems (usually LLMs) can assume several different roles and target different aspects of the collaboration. The ultimate aim of supporting groups is to afford interaction patterns that benefit groups in terms of achieving their goals, for instance learning. Learning may encompass the acquisition of domain- specific knowledge as well as knowledge that enables group members to collaborate more effectively. It is important to note that beneficial effects of a support tool depend on whether and to what extent the tool can elicit conducive interaction patterns. Against this background, asking about “the” effect of generative AI on collaborative learning is not precise enough. Instead, we should wonder how learners and groups interact with a tool while working on a task, which individual and collaborative learning processes are sparked in this way, and how this in turn affects outcomes such as the co-construction and acquisition of knowledge. This insight comes from the so-called media/methods debate which is essential for contemporary research on technology-enhanced leaning (see Clark, 1994; Kozma, 1994; Tennyson, 1994; Weidlich et al., 2025; see Sailer et al., 2024 for a meta-analysis). In the following, we briefly summarise the evidence on the effectiveness of generative AI support for learning.

Of the studies described above, several employed a quasi-experimental or experimental design, usually in the field (such as university courses), which is the gold-standard to investigate effects of support on outcomes such as domain-specific knowledge (Lin et al., 2024; Naik et al., 2024; Naik et al., 2024). Notably, the studies we identified all focused on knowledge about computer programming. For example, Lin et al. (2024) conducted a field experiment with university students, comparing the effect of support from a GenAI system on knowledge about Arduino programming with a control condition that did not receive any support during collaboration. The GenAI system served as a chatbot that provided information in the first phase of the collaboration and later provided the group with a summary of the group’s current understanding as well as with topics the group should subsequently explore. The authors reported a medium-sized positive effect (η² = 0.035), suggesting that the groups benefited from the interaction with the chatbot of support in terms of knowledge about programming.

Naik et al. (2024) examined the effect of personalised reflection episodes in which groups of approximately three students were prompted to discuss specific aspects of their current tasks compared to a control condition without support. Using ChatGPT, the reflection prompts were tailored to content of the groups’ discussion. While the authors reported that all students increased their knowledge about programming, they did not find a statistically significant effect of the personalised reflection prompts on students’ knowledge.

In another study, Naik et al. (2024) investigated the effect of different sources of contrasting cases (self-generated, personalised by generative AI) on students’ domain-specific knowledge in computer programming. The authors compared a control condition that did not perform a contrasting cases comparison (but did perform a reflection activity) with a condition in which groups had to generate a contrasting case for their solution on their own, and with a condition in which ChatGPT was prompted to construct a contrasting case for the group based on the group’s solution. While the analyses did not yield an overall effect of the machine-generated contrasting cases on students’ knowledge, the results suggested that students with lower prior knowledge benefited more from machine-generated contrasting cases than did students with higher prior knowledge. In turn, students with higher prior knowledge showed a higher post-test score when their group had to generate a contrasting case on their own.

In summary, several studies investigated the effectiveness of different approaches regarding how systems that leverage generative AI support the acquisition of domain-specific knowledge. These studies tended to find positive effects of support on students’ knowledge. Nevertheless, it should be noted that the studies employed different approaches to fostering collaborative learning, and that due to methodological limitations, caution is warranted when interpreting the results. We will discuss these points in more detail at the end of this chapter.

To initiate and maintain effective collaboration that eventually leads to learning, groups need to perform a broad variety of actions (see e.g. King, 2007; Schürmann et al., 2023; Strauß et al., 2025). Despite the crucial role of collaboration skills, only a few of the studies we found investigated them as an outcome (or dependent) variable. One example is the study by Darmawansah et al. (2025), who compared argumentative speaking performance, as well as the complexity of the arguments, before and after collaborating with support from a generative AI system when learning English as a foreign language. In one phase, the groups could query ChatGPT for information about the topic, and in a subsequent phase, groups used predefined prompts for ChatGPT that processed the input from the group, for instance "Rewrite these arguments using the argumentation model”. The results revealed a significant large effect (η² = 0.33) of the scaffolding on students’ argumentative speaking performance and on the complexity of the arguments that students were able to provide after the collaboration.

Another study, by Wei et al. (2025), examined the effect of a combination of different GenAI tools (ChatGPT, Midjourney, Runway) during a digital storytelling activity on students’ collaborative problem- solving skills. Over several weeks, students created digital videos that told short stories. Groups in the experimental condition used the different GenAI tools to create storyboards, images, and videos. As the dependent variable, the authors used a questionnaire on which participants rated statements regarding their past collaborative behaviour and potential reactions to hypothetical collaborative situations. Learners who collaborated using these tools achieved a higher score on this measure than did their peers who collaborated without GenAI tools (large effect, η² = 0.16).

Taken together, there is still very limited research on the effects of different approaches to using generative AI in order to scaffold collaborative learning with the aim of promoting collaboration skills. Further studies in the field have investigated other outcomes or characteristics of the collaboration process, such as group performance (An et al., 2024; Naik et al., 2024), interaction patterns and participant roles that emerged during collaboration when participants had access to a generative AI system (An et al., 2024; Gong et al., 2024; Zhu et al., 2023), reflective thinking (Lin et al., 2024), the perceived influence of ChatGPT on group dynamics (dal Ponte, Dushyanthen and Lyons, 2023) or the overall perception of a generative AI chatbot (Cai et al., 2024).

Rationales guiding the adaptivity

Research emphasises that scaffolds are effective if they are designed in accordance with didactic principles (Kollar et al., 2025). Against this background, we explored how generative AI systems were designed so as to provide groups with adaptive support.

One approach utilised by authors to afford personalised scaffolding was to rely on the generative model (i.e. LLMs) to perform monitoring, diagnosis, and decisions about specific instructional actions (e.g. Chu, Xu and Zhai, 2024; Greisel et al., 2025; Jong et al., 2024). In some cases, groups could use the generative AI as they wished; consequently, the specific adaptive actions from the system were dependent on groups’ prompts or the behaviour that was transmitted to the LLM (e.g. Chu, Xu and Zha, 2024; contributions in Greisel et al.,2025; contributions in Hernandez-Leo et al., 2025; Wei et al., 2024). In these cases, if the model received a request aimed at receiving feedback or guidance (e.g. Gong et al., 2024; Suciati et al., 2024), the guidance would be dependent on the qualities of the model (i.e. LLMs such as ChatGPT 3.5).

Other researchers used prompts to instruct the GenAI system how it should react. One such approach is to specify a role that the system should assume and specific evaluation criteria that the system should use to assess inputs from a group (e.g. An et al., 2024; Dang et al., 2025; dal Ponte, Dushyanthen and Lyons, 2023; Feng, 2025). For instance, groups participating in the study by dal Ponte et al. (2023) were tasked to develop evaluation plans, which were subsequently assessed by ChatGPT. The authors provided the model with an instructional prompt delineating the role that the model should assume (i.e. an expert) as well as criteria against which the model should evaluate the group's solution: “Act as an expert in the learning health systems framework with a focus on socio-technical evaluation plans. [...] Meticulously analyse its content based on the specific categories” (p. 2). These categories included the method of evaluation: “Assess the appropriateness, robustness, and feasibility of the chosen evaluation method” (p. 2), or the data source: “Critically examine the listed data sources for their relevance, reliability, and potential to address the evaluation’s objectives” (p. 2).

An example of a prompt that was used to design an artificial peer can be found in An et al. (2024), who tested a chatbot for collaborative learning. The authors gave GenAI models (Ernie-Speed-128k and Qwen LLM) prompts such as “You are a student agent named Alice who will participate in a group discussion with several students on [topic]” (p. 3), or “Once the discussion starts, you should guide the conversation and help students delve deeper into the questions being discussed by sharing various viewpoints.” (p. 3). The artificial agent was implemented in a chatroom where students could collaborate and participated in the discussion.

A more complex system, which makes use of multiple LLM-based agents, was described by Wu et al. (2025), who examined a multi-agent system that was used to support groups of learners during a programming task. The authors specified the system’s behaviour by giving it directions about the behaviour the system should exhibit (e.g. style and persona), the format the system’s responses should have, and example responses (Wu et al., 2025).

An alternative approach to researchers or teachers providing the evaluation criteria that should be used by a generative AI system is to allow the groups to determine the criteria themselves. One such example was presented by Ko and Foltz (contribution in Greisel et al., 2025), who described a generative AI system that monitors groups’ discussions against the background of group norms that were designed by the groups prior to the collaboration. Afterwards, the system facilitates the reflection on the collaboration by providing examples from the collaboration process in which these norms were visible.

As an alternative to using qualitative instructions to steer adaptive interventions by the generative AI system, interventions may be based on production rules (IF-THEN statements; some authors also use the term “triggers”, e.g. Naik et al., 2024). For instance, the system described by Cai et al. (2024) counted the number of contributions made by each group member. If the relative number of contributions of a group member fell below 10%, then the system would issue a prompt to encourage this group member to participate more. Similarly, Naik et al. (2024) implemented several production rules that used expression-based matching to identify specific types of code fragments that groups used while solving their programming tasks. Each production rule was aimed at encouraging reflection about one of five learning goals. For example, if the code entered by the students included a specific expression that aimed at altering the data type of a column, then the system would prompt the group to discuss their selection.

Two further approaches to guide a generative AI system in providing instructions to groups are to fine-tune the system and to provide it with additional information about the learning context. Fine-tuning consists of adapting a pre-trained model for specific tasks or use cases using a smaller, specific dataset. One example of fine-tuning can be found in An et al. (2024). While the authors stated that they used “extensive tutor dialogue data [...] to train the LLM, ensuring the generation of more professional responses for instructional and pedagogical guidance” (p. 2), they did not provide further details regarding the corpus of tutoring dialogue or the training procedure. Lin et al. (2024) also mentioned that they fine-tuned ChatGPT before implementing it into their GPT-Assisted Summarization Aid (GASA), but did not provide further information on this.

Instead of fine-tuning an LLM, Feng (2025) provided their LLM chatbot with presentation slides and a description of the collaborative problem-solving task that were used in the course that students attended, in addition to specifying the chatbot’s role and desired behaviour (see above). Feng (2025) argued that this approach is associated with lower costs than fine-tuning and allows for adjustments as soon as the system is in use.