Conference Agenda

Introduction.

Socio-emotional relationships between learners and educators are crucial for successful and joyful learning. Prior research (Hattie, 2023; Snijders et al., 2020) highlights cognitive and affective facets of effective teaching-learning interactions. These include the expression of trust, emotional warmth, responsiveness to learners’ individual prerequisites, instructional clarity. Didactic approaches such as Socratic dialogue build on these assumptions (Degen, 2025).

New advances in AI make it possible to implement such didactic models in intelligent learning environments. These systems enable reciprocal, naturally sounding verbal interaction between an AI tutor and a learner. However, it remains unclear how this communication is perceived by learners. To what extent can a positive socio-emotional relationship be established, and can learning and satisfaction be enhanced? These questions formed the core of the present study which compared student learning with an LLM-based AI versus a human tutor.

Two concepts on socio-emotional experience in media-based communication provided the theoretical foundation: social presence (SP; Biocca et al., 2003) and parasocial relationships (PSR; Sheng et al., 2025). SP refers to the subjective sense of social connection through reciprocal interaction. It is a multidimensional construct (Wang et al., 2025) encompassing cognitive and affective dimensions that support mutual understanding between communication partners. PSR describes the formation of an imagined relationship between a recipient and a media persona in one-sided communication contexts. However, the concept has been expanded to include mutual interaction scenarios in online-learning settings and communication with virtual entities such as robots or AI agents (Sheng et al., 2025).

Method.

An experimental design was used in which psychology students were randomly assigned to one of three conditions: learning with (1) a Socratic ChatGPT tutor via written chat, (2) a human Socratic tutor via written chat, or (3) a human Socratic tutor via Teams voice chat. Participants knew whether they interacted with an AI or a human tutor. Data were collected at three time points: (t1) a pre-study survey on demographic information and a prior-knowledge test, (t2) a lab-based learning session, (t3) a post-session survey including another performance test and measures of learning experiences. Learning content focused on applying APA standards in scientific writing, and explanations were standardized across conditions to ensure consistency.

At t3, the following variables were measured (all 5-point Likert scales):

- SP between learners and the AI or human tutor (Biocca et al., 2003), including three cognitive dimensions (item examples: “I tended to ignore ChatGPT [the tutor])”, “I was able to understand what ChatGPT [the tutor] meant”) and the affective dimension of co-presence (“ChatGPT [the tutor] influenced my mood while learning“).

- PSR towards the tutor (e.g., “I felt trust toward ChatGPT [the tutor]”; Paechter et al., 2000).

- Satisfaction with learning and teaching (6 items, e.g., “I learned a lot from ChatGPT [the tutor]”; Rindermann, 2009).

Results and Discussion.

Overall, the instruction was successful: in all groups students improved their knowledge from t1 to t3. However, no differences in performance were found across groups. Similarly, a MANOVA with experimental group and gender as independent variables revealed no group differences in the cognitive dimensions of SP. However, co-presence and satisfaction were higher in group 3 than in group 1. In addition, learning with a human tutor (group 2 or 3) lead to higher PSR. None of the analyses revealed significant effects for gender, or its interaction with group.

Because the explanations provided by the AI and human tutors were standardized, the absence of group differences in the cognitive dimensions of SP or in performance was expected. AI was therefore as effective as the human tutor conditions in supporting learning outcomes. Indeed, AI tutors offer advantages such as delivering accurate explanations, immediate feedback, and individualized support, which can foster cognitive achievement. Differences emerged in the socio-emotional variables, favoring the human tutor conditions. These variables are crucial for sustained effort and motivation.

Educational Significance of the Research.

Altogether it seems that while AI tutors can effectively support cognitive learning processes, they may not fully meet learners’ emotional needs during instruction. These findings have important implications for instructors and learners, especially regarding satisfaction and learning outcomes. They indicate that involving human tutors may be particularly beneficial at certain stages of learning or for key socio-emotional functions such as motivation and feedback.

In this positioning paper, we ask how large language models can enhance students’ learning when they write. When Norwegian universities advise students on their use, large language models are referred to both as tools and as learning partners (University of Oslo, 2025). “Tool” is a well-established term for digital writing support, but “learning partner” arrived with the large language models. The resource is humanized, much like when ChatGPT is described as a super-energetic colleague who, unfortunately, is also a compulsive liar (Strümke, 2023). This is, after all, how the language models appear. Although they have no consciousness, we read consciousness into the texts they produce.

ChatGPT and learning

Two cognitive models of the writing process form the starting point for examining what ChatGPT-supported writing means for students’ knowledge development and learning. Flower and Hayes’s “Cognitive Writing Process Model” (1981) is used as a starting point to discuss what the use of large language models means for the organization of writing, while Galbraith’s “dual process model” provides the basis for asking what large language models mean for writing as a process of discovery.

In Flower and Hayes’s model, the writing process is broken down into three parts: planning, writing (translation), and revision. Flower and Hayes argue that help in organizing the relationship among these three parts can reduce the cognitive load on the writer. They later revised the model several times, but this tripartite structure has been retained, both by Flower and Hayes and by a number of others who have developed models of writing as a cognitive process (see Becker, 2006). The model provides a framework for examining the role ChatGPT plays in organizing the writing process.

Galbraith (2009, 2018) emphasizes that it is the act of writing itself—and not the organization of writing—that strengthens learning and reflection. He argues that Flower and Hayes’s model, like the models developed somewhat later by Bereiter and Scardamalia (1987), “overemphasize the importance of explicit thinking processes in writing and hence treat production processes as a relatively passive component of the writing process” (2009, p. 24). Galbraith therefore proposes “a dual-process model of writing” in which the retrieval of knowledge is a consequence of the rhetorical writing task to be solved. This knowledge must be processed in writing (knowledge constitution) in order to become a response to the task (2009, p. 21). Writing thus evokes a conflict or tension between the knowledge that is retrieved and what the task demands. It is this conflict that generates a deeper understanding of what one is writing about. Therefore, according to Galbraith, writing is “intrinsically a process of discovery” (2009, p. 22). It is only when you write that you see what you think, and thus also “discover” how you can understand what you are writing about.

How can writing with ChatGPT enhance student's learning?

First, students can give ChatGPT instructions that organize their writing. If students instruct ChatGPT to take over the parts of writing that they already master rhetorically—for example, outlining, searching for sources, language editing, or translation—they can free up effort for other parts of the writing process that contribute more to learning and knowledge development. In that case, they must have the skills to ensure that quality control of ChatGPT’s contributions is less labor-intensive than doing these parts of the writing themselves.

Second, they can give ChatGPT instructions that strengthen writing as a process of discovery. Contributions from ChatGPT in the form of more knowledge and/or more rhetorical knowledge and more text help to calibrate the conflict between knowledge constitution and the rhetorical processing of knowledge during writing that Galbraith points out. ChatGPT can thus strengthen knowledge development by creating a more productive foundation for “discovery” in the writing process.

Further discussion

In the paper, we will examine more closely the prerequisites for students’ disciplinary understanding and reflection to be strengthened—and not weakened—when ChatGPT supports their writing. Finally, we ask what large language models such as ChatGPT will mean for educational institutions’ dependence on writing as a technology.

Educational research, especially educational technology research, has long grappled with a persistent paradox: introducing identical tools, curricula, and resources into different classrooms rarely yields the same results. Traditional theoretical models often attempt to explain this by categorising the static conditions required for success, such as specific teacher knowledge domains or levels of resource access. However, we argue that these approaches suffer from a fundamental theoretical deficit because they map the educational context without explaining the mechanism of its unfolding. By treating integration as a fixed state rather than an ongoing, developmental trajectory (Lachner, Backfish & Franke, 2024), they tell us what conditions are present but fail to explain how teachers translate those conditions into lived practice. To address this limitation, we introduce the Socio-Ecological Processual Enactment (SEPE) model. This theoretical framework, built abductively through an iterative dialogue between surprising empirical variations in practice and ecological theory, shifts the analytical focus from static competence to dynamic enactment.

The SEPE model is grounded in the recognition that teaching is not merely techne, defined as the productive craft of following a blueprint to create a known end, but rather phronesis, or practical wisdom exercised in conditions of uncertainty (Bardone, Mõttus & Eradze, 2024). Drawing on Aristotelian ethics and contemporary scholarship on situated cognition (see e.g., Hickey & Riddle, 2025), we argue that pedagogy is an emergent property of enactment rather than a fixed attribute of a teacher or a tool. In this view, educational tools, whether digital technologies, textbooks, or classroom arrangements, do not possess inherent pedagogical value. Instead, they are abstract artefacts that only gain meaning through the process of concretization within a specific context (Dron, 2022). This process relies heavily on the teacher's phronesis, which is the capacity to make situated instructional judgments in real time. The SEPE model posits that meaningful pedagogy arises not from compliance with external standards but from the ongoing interpretation of the learner ecology and the negotiation of competing goals.

Building on Bronfenbrenner’s Ecological Systems Theory and Process-Person-Context-Time framework (see Bronfenbrenner & Morris, 2006), SEPE structures the educational environment into nested layers with explicitly including a dynamic Process Layer that mediates between the person and the context.

• The Person Layer reframes the educator not just as a possessor of skills but as an agent defined by Attributes (e.g., subject role), Capacities, and Dispositions. While capacities provide the potential for action, the teacher's disposition channels how they choose to act.

• The Process Layer captures the proximal processes, which are the reciprocal and bidirectional interactions between teacher and student. It is here that phronesis is operationalised through pedagogical reasoning and action, not as a direct output of competence but a situated negotiation in which competing operational, motivational, and cognitive goals are dynamically balanced against the immediate realities of the learner ecology.

• The Context Layers situate these processes within nested environments: the Microsystem (the immediate instructional environment of activities, roles, and relations), the Mesosystem (linkages between the classroom and other settings like professional communities), the Exosystem (institutional decisions and resource allocation), and the Macrosystem (broader societal values, policies, and infrastructure).

• The Chronosystem argues that pedagogy is inherently temporal. It distinguishes between micro time, or the moment-to-moment instructional pivots required by immediate feedback, meso time (the sedimentation of routines), and macro time, or the influence of socio-historical shifts.

This lens reveals that expert pedagogy is often a matter of tinkering, defined as an iterative process of adjustment over time, rather than the immediate execution of a perfect design. The SEPE model offers a necessary corrective to the black box of educational research. By centring enactment and phronesis, it challenges the field to move beyond measuring simply in- and outputs. Instead, it frames teaching as an entangled, socio-material practice in which context and agency are inextricably linked. Theoretically, this implies that we cannot understand pedagogical success by looking at the teacher or the tool in isolation. We must look at the negotiated process through which they mutually constitute one another. Practically, teacher development should focus less on technical mastery of tools and more on cultivating adaptive expertise. This requires the phronetic capacity to interpret evidence and to orchestrate complex learning environments in response to students' emergent needs.