Students' conceptions about AI in science classrooms

Abstract

As Artificial Intelligence (AI) becomes increasingly integrated into school contexts, debates often focus on efficiency, trustworthiness, automation, and concerns about students' growing reliance on such systems. However, less is known about how students themselves conceptualise “an AI for school” and which pedagogical roles they attribute to it. This study investigates students' expectation profiles regarding educational AI, the pedagogical meanings they attach to AI functions, and how they articulate the balance between AI support and cognitive independence. Drawing on 841 open-ended responses from students across multiple European countries and languages, the study employed an exploratory mixed-methods design using a dual coding framework. Functional expectations (System A) were analytically differentiated from pedagogical meanings and values (System B), followed by quantitative co-occurrence analysis and residual mapping. The findings reveal systematic associations between expected AI functions and pedagogical meanings, resulting in three recurrent role repertoires: Tutor-oriented expectations emphasising explanation and understanding; Coach-oriented expectations focusing on personalisation, organisation, and self-regulation; and Companion-oriented expectations foregrounding relational, emotional, and ethical dimensions. Across all profiles, students articulated a clear normative boundary between supportive guidance and inappropriate delegation, favouring AI that scaffolds thinking during problem-solving rather than replacing cognitive effort. These findings contribute to digital education research by demonstrating that the educational relevance of AI depends less on technical functionality alone than on how AI affordances are pedagogically framed and aligned with learner agency, cognitive autonomy, and responsible classroom integration.

1 Introduction

Public debates on Artificial Intelligence (AI) in education are often dominated by concerns about learner passivity, over-reliance, and the outsourcing of cognitive work. As AI increasingly enters school contexts, it is often framed as a tool for efficiency, automation, or individualised instruction. At the same time, research on responsible and human-centred AI highlights that the educational implications of AI extend beyond performance gains, encompassing issues of fairness, transparency, privacy, and, most critically, learner agency and autonomy (Fu and Weng, 2024). Understanding how students conceptualise AI in educational settings is, therefore, a key requirement for didactically sound and ethically responsible integration.

AI is increasingly framed as a complementary resource that can support learning processes oriented toward deep understanding, self-regulation, and meaningful learner engagement (Fu and Weng, 2024; Lan and Zhou, 2025). This aligns with evidence that inquiry-based learning environments can foster intrinsic motivation and active engagement when learners have autonomy and space for creative problem solving (Conradty and Bogner, 2019). From this perspective, the educational relevance of AI lies not primarily in its capacity for automation. Of interest is which technical affordances of AI are deliberately aligned with pedagogical intentions and embedded within classroom practices. This shift foregrounds a critical concern: how AI might support learning while preserving, rather than displacing, students' cognitive agency and responsibility.

This perspective is reflected in the Deeper Learning Competence Framework of the Erasmus + project labelled Discovery Space (Sotiriou and Bogner, 2023), which conceptualises learning as an active sense-making process centred on learner agency, self-regulation, collaboration, and ethical awareness. Within this framework, AI-supported learning environments are understood as pedagogical scaffolds that support the development of problem-solving competence. Examining learners' own conceptions of what AI should do and what purposes it should serve thus provides an empirical entry point into the alignment between emerging AI practices and deeper learning goals. AI environments, designed to continuously monitor student progress, provide targeted feedback and assess the student's mastery. All this information might be collated throughout a student's time in formal (and, in some cases, informal) educational settings.

Recent research on students' perceptions of generative AI presents a nuanced picture. While students express interest in AI as a potential support for learning, persistent concerns remain regarding reliability, ethical implications, and the risk of over-reliance (Chan and Hu, 2023; Zhai et al., 2024). These findings highlight a core didactic tension: whether AI should function primarily as a provider of answers or as a scaffold that supports thinking and learning without displacing learners' cognitive work. This tension is further intensified when AI is conceptualised not merely as a tool, but as a learning companion capable of adopting multiple roles within the learning process. Research on educational chatbots and AI companions suggests that learner engagement and trust are strongly influenced by how such systems are experienced in interaction, including their relational and socio-emotional dimensions (Maples et al., 2025; Mills et al., 2025). However, large-scale empirical evidence capturing students' expectations and role conceptions across languages and educational contexts remains limited.

A further gap concerns students' mental models of AI. Research shows that learners' conceptions of how AI works, what it can be trusted with, and what counts as appropriate help vary systematically and have direct implications for AI literacy and classroom scaffolding (Dangol et al., 2025). These insights reinforce the need to study not only which AI functions students expect, but also the educational meanings and values they attach to these functions. To address these gaps, the present study (conducted within the DISCOVERY SPACE initiative) investigates students' conceptions of “the optimal use of AI in schools” using open-ended responses elicited before an engagement with inquiry-based future-oriented learning scenarios during experimentation with virtual online labs. An open-question format was chosen to capture learners' own ways of articulating expectations, assumptions, and concerns, without imposing predefined roles, functions, or evaluative frames. Such an approach is well established in research on learners' conceptions and mental models, where open elicitation is considered essential for accessing underlying meanings rather than surface preferences (Dangol et al., 2025). Studies on learners' conceptual frames show that open responses often form coherent interpretative patterns rather than isolated statements. This enables reconstruction of role-related expectations (Conradty and Bogner, 2020).

In the DISCOVERY SPACE intervention, the formative assessment process was fully embedded in the problem-solving practice. For this reason, an Exploratory Learning Environment was used to facilitate the introduction of problem-solving practices that support learning and problem-solving using online experiments. This environment was enabled with an AI-driven learning companion to provide students with the necessary support and guidance, and with advanced web-based interfaces to enhance the learning experience and facilitate collaboration, modelling, and problem-solving. While using the system, students were encouraged to actively construct their own knowledge by exploring the learning environment and making connections with their existing knowledge schema. The role of AI was to minimise the cognitive overload that is often associated with exploratory learning (Sotiriou et al., 2020) by providing automated guidance and feedback, based on knowledge tracing and machine learning. This feedback addresses misconceptions and proposes alternative approaches to support students as they perform the problem-solving tasks. The AI-driven learning companion supported students to develop their own learning paths within a self-regulated learning process that provided structure to the student experience, allowing each learner to pursue goals that require extended engagement or persistence across multiple contexts and learning opportunities.

In this study, closely related constructs - learner agency, cognitive autonomy, and self-regulation - are analytically distinguished but treated as interconnected dimensions of learner engagement. Learner agency is an overarching concept describing the capacity of learners to act intentionally and make meaningful choices in learning contexts. Cognitive autonomy refers more specifically to independent thinking and understanding, while self-regulation captures processes of planning, monitoring, and managing one's own learning. Algorithmic guidance refers to structured support provided by AI systems, that may shape but do not replace learner agency. This distinction allows us to examine how students link their expectations of AI support to different dimensions of active learning.

The objectives of our study were threefold. First, we examined what expectation profiles and role repertoires students articulate for an AI learning companion and how these expectations are conceptually organised (RQ1). Rather than assuming predefined categories, this objective focused on understanding how learners themselves describe AI and how their expectations are conceptually organised across different dimensions of use and meaning.

Second, we analysed how functional requirements are pedagogically framed through educational meanings and values (RQ2). In particular, we investigated how functional requirements, such as explanation, feedback, or personalisation, are interpreted and qualified through educational attributions of meaning, including independence, self-efficacy, relational support, and fairness. This objective addresses the translation from functional descriptions of what AI should do to didactic principles that define why and how these functions are considered educationally valuable.

Third, we addressed a central normative tension in current debates on AI in education: the balance between assistance and autonomy (RQ3). We analysed how students articulate the distinction between AI as a provider of solutions and AI as a support for thinking and learning, and how they define the boundaries of legitimate help. This objective sheds light on learners' perspectives on agency, responsibility, and the role of AI in preserving meaningful learning activity.

2 Methodology

2.1 Study design and data collection

The study is based on the analysis of students' open-ended responses to a single exploratory prompt administered after engagement with a set of future-oriented learning scenarios. In total, N = 841 written responses were collected through a digital learning environment designed for classroom use (https://trust-ai-lab.eu/). The platform is primarily used in secondary school contexts across several countries. Responses were provided in English, Greek, Portuguese, and Spanish. Participation was voluntary and anonymous.

In line with data protection regulations (GDPR) and the ethical framework of the project, the research team did not have access to identifiable student information. The dataset, therefore, contains only textual responses and no demographic variables such as age, gender, or school affiliation.

Based on the design of the learning environment and the participating school contexts, users are assumed to be students approximately 11–18 years old, with the majority likely to be in lower secondary education (around grades 7–8). However, no precise demographic breakdown of participants was available to the researchers. Students were asked the following open-ended question: “Imagine you can invent a perfect AI for school - what should it be able to do?” The open format was chosen to elicit students' spontaneous conceptions without constraining responses through predefined categories or scales.

Given the scale of the dataset, the study combines qualitative coding with AI-supported analytical procedures to support the systematic organisation and comparison of student responses. Recent methodological work highlights that large language models can assist in scaling qualitative analysis, such as categorisation, pattern detection, and cross-case comparison, when used transparently and in conjunction with human oversight, rather than as autonomous analytic agents (de Paoli, 2024; Hayes, 2025). In this sense, AI is employed not as a substitute for interpretation, but as a methodological support for analysing students' conceptions on a scale while preserving interpretive sensitivity.

The present study employs an exploratory mixed-methods design, integrating qualitative interpretation with structured category analysis. The objective is not to infer causal relationships, but rather to identify recurring patterns in how students conceptualise the potential roles and functions of AI in school learning contexts.

2.2 Translation and data preparation

All responses were translated into UK English prior to analysis. Given the brevity and syntactic simplicity of most statements, full back-translation procedures were not applied. Instead, translations were checked for consistency of meaning across languages and revised only in cases of clear semantic ambiguity. While full back-translation was not applied, particular attention was paid to preserving the functional meaning of responses, which typically consisted of short and direct statements about desired AI capabilities. Each response was treated as a single coding unit. Where responses contained multiple ideas, all relevant aspects were coded within the same unit.

2.3 Analytical framework and coding procedure

Coding was conducted using a combined human-AI approach. Large language models were employed to facilitate the organisation and preliminary categorisation of responses. The model-assisted coding was implemented under controlled prompting conditions to generate candidate category assignments, which were subsequently reviewed and consolidated within the established coding framework. The coding assistance was provided by a GPT-4 class language model via the ChatGPT interface.

As is common in both manual and AI-assisted qualitative coding, the category definitions were refined iteratively during the coding process before proceeding to the final stages of analysis. The development of the coding scheme, including the categories and their definitions, followed established qualitative research practices.

To assess the consistency of the category assignments, a human coder reviewed a randomly selected subset comprising 50% of the responses. This step served to identify ambiguous cases and evaluate the coherence of the coding scheme. Where discrepancies occurred, their underlying causes were analysed, and the category definitions were further refined to improve conceptual clarity. This iterative refinement process ensured a high level of consistency and quality in the final coding.

Inter-coder reliability was assessed for a subsample of 100 statements using Cohen's Kappa, indicating substantial agreement across categories (κ = .891). This procedure ensured consistent application of both category systems across the dataset.

2.4 Workflow with LLM

To support the systematic analysis of the large corpus of responses, a large language model (LLM) was used in a multi-step, research-guided procedure. The coding schemes were not derived inductively from the responses alone but were based deductively on relevant literature and the theoretical framework of the study. Prompts can be found in the Appendix.

In a first step, the researchers developed preliminary categories based on the literature, representing potential dimensions of responses to the open-ended question. In addition, the LLM was asked to suggest categories that might appear in responses to the prompt. These suggestions were compared with the literature-based categories.

Based on this comparison, a consolidated coding scheme was created. The scheme was then provided to the LLM with the task of refining the wording of the categories, formulating short operational definitions, and suggesting illustrative examples. This output was reviewed and, where necessary, revised by the researchers.

In the next step, the dataset was provided to the LLM in structured form (ID; text response) to generate an initial coding of the responses using the preliminary coding scheme. This first coding was reviewed by the researchers to identify ambiguous definitions, conceptual overlaps, or problematic assignments. Where necessary, category definitions were further refined and illustrated with quotations from the dataset.

The final coding of the complete dataset was then conducted using the revised coding scheme. Multiple coding was allowed. Missing responses were coded as “0”, and non-classifiable responses would have been coded as “–1”. In the present dataset, however, all responses could be assigned to at least one category. The output was generated in structured form, including the response ID and the assigned category or categories.

This procedure was conducted separately for both coding schemes. Finally, the codings from System A and System B were merged using the response IDs. These merged data were used to analyse the frequency and co-occurrence of categories across both systems.

2.5 Quantitative analysis of category distributions

Category frequencies were calculated separately for System A and System B to identify dominant functional expectations and pedagogical meanings. These distributions served as a descriptive basis for further analysis and are reported in Figure 1.

Figure 1

2.6 Co-occurrence analysis and statistical testing

To explore associations between functional requirements and pedagogical meanings, a cross-tabulation (A × B matrix) was constructed linking all System A and System B codes at the response level (Figure 2), representing the joint distribution of functional and pedagogical categories.

Figure 2

A chi-square exploratory analysis of association patterns was conducted to assess whether functional and pedagogical categories were associated beyond chance. Effect size was estimated using Cramer's V. To identify specific over- and under-represented category combinations, standardised residuals were inspected and visualised in a residual heatmap (Figure 3).

Figure 3

As individual responses may receive multiple codes within each category system, the resulting A × B combinations are not statistically independent. In line with established approaches to the analysis of multiple-response and code-based data (e.g., Mahieu et al., 2021), the chi-square statistics, standardised residuals, and Cramer's V are therefore interpreted as exploratory indicators of association patterns rather than as strict tests of independence.

This interpretation is consistent with recent mixed-methods studies employing code-based content analysis, where categorical associations are analysed to identify structured patterns while acknowledging the limitations of independence assumptions (e.g., Schiltz et al., 2024; Waters et al., 2024). Accordingly, the statistical results are used to support the identification of meaningful co-occurrence structures, rather than to infer population-level relationships.

2.7 Identification of expectation profiles and role patterns

Based on frequency distributions and residual patterns in the A × B co-occurrence matrix, recurrent constellations of category combinations were identified across responses. These constellations reflect systematic associations between functional expectations (System A) and pedagogical meanings (System B).

The identification of expectation profiles followed explicit analytical criteria. A constellation of categories was considered indicative of a profile when it (1) occurred with notable frequency across the dataset, (2) showed distinctive over-representation in the co-occurrence analysis, and (3) formed a coherent interpretative pattern in terms of how students conceptualised the role of AI in educational contexts.

The interpretation of such constellations as role-oriented expectation profiles is consistent with prior research in the field of artificial intelligence in education, which conceptualises educational technologies and agents not as uniform systems but as entities fulfilling distinct pedagogical roles (Baylor and Kim, 2005; Schiff, 2022). In particular, the differentiation between roles such as tutor, coach, and learning companion has been widely used to describe variations in instructional support, guidance, and socio-emotional interaction in AI-supported learning environments (Baylor and Kim, 2005; Johnson et al., 2000).

The resulting profiles should therefore be understood as descriptively and analytically derived patterns, rather than statistically inferred latent types. Based on these criteria, three dominant profiles were identified and labelled as Tutor-oriented, Coach-oriented, and Companion-oriented, reflecting established distinctions in the literature while remaining grounded in the empirical response patterns of the present dataset.

3 Results

The results are reported in relation to the three research questions and are based on the joint analysis of functional requirements (System A) and pedagogical meanings and values (System B).

3.1 Expectation profiles and role repertoires (RQ1).

Students' expectations (RQ 1) towards an AI learning companion are distributed across both functional requirements (System A) and pedagogical meanings (System B). Joint analysis of the A × B frequency matrix (Figure 2) and the Residual-Z-Heat Map (Figure 3) suggests the presence of three recurrent clusters of expectations within the dataset. These clusters are not homogeneous blocks but diagonal constellations of function-meaning alignments, indicating that AI roles are frequently described in relational combinations of functions and meanings rather than as isolated features.

Of interest are responses combining functional and pedagogical dimensions, illustrating specific cross-category pairings with high Z values (Table 3).

Given the multi-coded nature of the dataset and the resulting lack of independence between observations, the statistical measures (χ², residuals, and effect sizes) are interpreted as exploratory indicators of patterns within the dataset rather than as inferential evidence of generalisable associations.

3.2 Framing of functional requirements through pedagogical meanings (RQ2)

Both functional requirements (System A) and pedagogical meanings (System B) are widely represented. The A × B matrix (Figure 2) indicates patterned co-occurrences between specific functions and meanings.

Explanation and learning assistance (A1) most often co-occur with knowledge and understanding (B1). Practice and feedback (A2) are mainly linked to thinking skills and independence (B2). Personalisation (A3) is most frequently associated with self-efficacy (B3), while organisational support (A4) is primarily linked to responsibility (B4). These pairings are reflected as comparatively higher residual values in the heatmap (Figure 3), which should be interpreted as exploratory indications rather than statistically independent associations.

3.3 Help vs. independence (RQ3)

Functions related to explanation, practice, and feedback (A1, A2) are among the most frequent expectations (RQ 3), while explicitly delegating functions occur less often. At the same time, meanings related to understanding, thinking skills, and independence (B1, B2) are highly prevalent.

In the A × B matrix (Figure 2), learning assistance and feedback are most frequently combined with meanings related to understanding and independent thinking (A1 × B1, A2 × B2). Residual analysis (Figure 3) suggests that these combinations occur more frequently than others in the dataset, although these patterns should be interpreted cautiously given the lack of independence among observations.

4 Discussion

4.1 Help vs. independence (RQ3)

At the level of articulated expectations, many responses indicate a normative boundary between legitimate support and inappropriate delegation. Rather than expecting AI to replace learning activities, responses predominantly frame AI assistance as explanation, guidance, feedback, and orientation in order to support their role as active learners. This focus on guided support rather than substitution aligns with research showing that inquiry-oriented learning environments foster engagement and autonomy when learners stay cognitively active (Conradty and Bogner, 2019). This finding aligns with recent ethical and policy-oriented research, which emphasises that the educational risks of AI do not stem from the technology per se, but from unregulated and pedagogically unreflective use (Miao and Holmes, 2023; Williams, 2024). In consequence, a central contribution of the present study lies in challenging this one-sided narrative. The present findings suggest that fears of inevitable learner passivity may be overstated. When learners themselves articulate clear expectations regarding the limits of acceptable AI support, the pedagogical task shifts from prohibition to framing. Schools are therefore not confronted with the question of whether AI should be used, but under which conditions its use supports meaningful learning. This view is consistent with calls for human-centred and responsible AI in education, which stress the importance of pedagogical validation and contextual regulation (Fu and Weng, 2024; Kasneci et al., 2023).

4.2 Roles and expectation profiles for AI (RQ1)

Against this background, the analysis of students' expectations provides important insight into how AI is imagined within the pedagogical landscape. Rather than describing isolated technical features, responses suggest coherent expectation profiles that can be interpreted as role repertoires for an AI learning companion (Conradty and Bogner, 2020). The three identified profiles, Tutor, Coach, and Companion, represent distinct yet interconnected orientations towards AI support, rather than fixed system roles. These roles are not introduced as a new taxonomy, but as empirically grounded interpretations of students' expectations that resonate with existing work on AI learning companions and role-fluid support systems (Chou et al., 2024; Han et al., 2025).

This interpretation is supported by responses that combine functional expectations with pedagogical meanings across category systems. As shown in Table 3, students frequently articulate expectations that simultaneously address a functional dimension (System A) and a pedagogical or value-oriented dimension (System B), for example by linking source transparency to understanding (A8×B1-Tutor), emotional support to relational communication (A9×B5-Companion), personalisation to fairness (A3×B7-Coach), or motivational features to inclusion (A5×B7-Coach). These cross-category combinations provide an empirical basis for interpreting the responses in terms of role repertoire rather than isolated features.

Importantly, these roles are not conceived as fixed identities. Instead, the data suggest that role expectations vary depending on learning goals, tasks, and situations. Tutor-oriented expectations emphasise explanation, feedback, and conceptual clarification; Coach-oriented expectations focus on organisation, personalisation, and self-regulation; Companion-oriented expectations foreground relational, emotional, and ethical dimensions of learning. This value-oriented framing reflects current approaches to responsible AI in education, which place learner agency, fairness, and transparency at the centre (Fu and Weng, 2024; Miao and Holmes, 2023). This role flexibility mirrors findings from recent research on intelligent learning companions, which describes learner expectations as role-fluid rather than role-specific (Giannakos et al., 2025; Han et al., 2025).

Companion-oriented expectations highlight the need for clear pedagogical framing. AI-based relational and emotional support raises important pedagogical considerations; it touches on issues of trust, privacy, and dependency. Research on motivation suggests that relational framing can significantly enhance engagement and perceived support but only when it is carefully designed (Conradty and Bogner, 2024). Students' emphasis on fairness, transparency, and respectful communication points to the need for clear boundaries and teacher oversight. The Companion role is not described in the data as a substitute for human relationships in learning, but as a limited form of support embedded in instructional contexts. Its educational value depends on transparency, restraint, and clear responsibility - and on keeping human relationships at the centre of learning (Erduran and Levrini, 2024).

For school practice, these findings suggest that students do not expect AI to assume a single, universal function. Instead, they envision AI as a pedagogical resource whose role depends on instructional context. This has direct implications for teaching: the central didactic question is not whether AI should be used, but which role it should play in a given learning situation. Making these roles explicit can help teachers guide appropriate use and avoid both over-delegation and unreflective reliance on relational features (Gligorea et al., 2023) These findings also point to the potential of adaptive learning environments that respond to students’ needs in real time (Sajja et al., 2024). At the same time, challenges such as resource constraints, assessment design, and the balance between individual and collaborative learning remain important considerations (Bernacki et al., 2021).

4.3 From functions to didactic principles: why features are not enough (RQ2)

While students often articulate their expectations in functional terms (such as explanation, feedback, or personalisation), these functions do not stand alone. An analytical contribution of this study is the distinction between functional requirements (System A) and pedagogical meanings and values (System B). The moderate association between the two systems (Cramer's V = .193) indicates systematic co-occurrence rather than equivalence. This distinction becomes visible in students' responses, where functional expectations are frequently qualified through explicit pedagogical framing. For example, explanation is requested not as solution delivery, but to support independent thinking. Such patterns illustrate that functions acquire educational relevance only when embedded in values such as understanding, autonomy, self-efficacy, fairness, or relational security. Instructional design alone does not ensure educational value; its features become meaningful only when they foster learners' motivation and autonomy (Conradty and Bogner, 2018, 2019). The findings help illustrate how both category systems matter within this dataset. A function like explanation (A1) only gains meaning when it supports understanding and independent thinking (B1). Similarly, personalisation (A3) is not valued solely for technical optimisation, but as a path to fairness and inclusion (B7). Even relational support (A9) is framed by expectations of respectful communication and trust (B5). On their own, these functions or values remain partial. Taken together, they explain why some forms of AI support are welcomed while others are not. The combination of System A and System B shows not just what students want AI to do, but what they want it to mean for learning.

Theoretically, this layered structure aligns with research on mental models, which shows that learners often express normative orientations through descriptions of actions and tools (Carley and Palmquist, 1992). Methodologically, the partial overlap between functional and meaning-based categories should therefore not be interpreted as weak discriminant clarity. In educational contexts, actions and values are closely intertwined. Similar arguments are made in research on responsible and human-centred AI, which emphasises that educational technology must be evaluated not only by what it does, but by what it supports pedagogically (Fu and Weng, 2024). For instructional design, this finding has clear consequences. AI integration cannot be guided solely by feature lists. Instead, functional affordances must be deliberately aligned with pedagogical intentions. Teachers thus play a central role in translating technical possibilities into didactic principles. As AI tools increasingly become an integral part of scientific practices, the element of agency has implications for science education. Learning environments will need to adapt to AI as a “collaborator” in scientific practices, where students and teachers jointly engage with such tools in addressing scientific problems (Erduran and Levrini, 2024).

Concerns about learner passivity, over-reliance, and the outsourcing of cognitive work may exist but apparently may be overstated. The present study contributes to questioning this one-sided narrative. At the level of articulated expectations, students themselves draw a clear normative boundary between legitimate support and inappropriate delegation. Rather than expecting AI to replace learning activity, they predominantly frame AI assistance as explanation, guidance, feedback, and orientation - forms of support that preserve their role as active learners. This finding aligns with recent ethical and policy-oriented research, which emphasises that the educational risks of AI do not stem from the technology per se, but from unregulated and pedagogically unreflective use. From a didactic perspective, this suggests that fears of inevitable learner passivity may be overstated. When learners themselves articulate clear expectations regarding the limits of acceptable AI support, the pedagogical task shifts from prohibition to framing. Schools are therefore not confronted with the question of whether AI should be used, but rather with the question of under what conditions its use supports meaningful learning. This view aligns with calls for human-centred and responsible AI in education, which emphasise the importance of pedagogical validation and contextual regulation.

4.4 Implications for teaching and the professional role of teachers

Taken together, the findings suggest a shift in how AI challenges teaching practice - not by replacing teachers, but by redefining professional responsibility. If AI can provide explanation, feedback, and organisational support, the teacher's role increasingly lies in orchestration, framing, and judgment. This view is consistent with research on teacher-AI complementarity, which emphasises that AI is most effective when embedded in teacher-led pedagogical decision-making (Holstein et al., 2019). From this perspective, the help-independence dilemma is not a technical problem, but a pedagogical one. Learners expect support that enables thinking, not one that substitutes for it (Sampson et al., 2011). Many responses indicate clear boundaries between support and substitution. Still, a smaller group explicitly expects AI to take over tasks such as homework or the production of answers. These expectations should not be read as mere misuse. They point to a deeper problem of assessment. As long as school evaluation rewards correct outputs rather than understanding, delegating work to AI remains attractive.

In a learning environment where knowledge is permanently accessible, this logic becomes increasingly challenged (Broadfoot and Rockey, 2025). Students themselves link AI functions to pedagogical values such as understanding, autonomy, and fairness. They ask for explanations, not ready-made solutions. Assessment, therefore, needs to shift accordingly. Tasks that require reasoning, justification, or reflection on learning processes reduce the value of simple delegation. From this perspective, AI does not necessarily weaken assessment but may expose the limits of traditional grading focused on factual recall (Khlaif et al., 2025). Used well, AI can support a move towards competence-oriented assessment that aligns with students' own expectations of meaningful learning support.

Meeting this expectation requires teachers to establish clear rules, transparent role definitions, and task designs that make appropriate AI use visible and discussable. Recent work on AI literacy in teacher education similarly highlights the need to strengthen pedagogical judgment, ethical reflection, and contextual decision-making, rather than focusing solely on technical competence (Allen and Kendeou, 2024). In this sense, AI does not diminish the importance of teachers. On the contrary, it increases the demand for didactic clarity, professional reflection, and pedagogical leadership. Beyond this shift in professional responsibility, the findings also point to a concrete didactic application: AI-supported concept elicitation in everyday teaching.

Open questions can make students' conceptions visible and create conditions for conceptual change, because teaching responds to what learners actually think rather than to assumed deficits (Duit and Treagust, 2003; Pacaci et al., 2024). In many classrooms, however, questions are experienced as assessment rather than interest, which can make answering feel risky and inhibit genuine engagement (Nystrand et al., 2003). While teachers can value open questioning, they usually lack the time to analyse patterns across many open responses, even though such interpretation is central to formative assessment (Veugen et al., 2021). Recent work shows that large language models can support the scalable analysis of open-ended responses without replacing pedagogical judgment (Hayes, 2025). In this sense, the present study illustrates not only what students expect from AI but also how AI can help teachers treat student answers as a resource for responsive teaching rather than as isolated indicators of success or failure.

4.5 Limitations

This study has several limitations that should be considered when interpreting the findings.

First, no detailed demographic information was available. Due to data protection regulations, only anonymised textual responses could be analysed, without metadata on age, gender, or school background. While the learning environments are typically used by students aged approximately 11–18, the findings reflect general patterns rather than age-specific differences.

Second, the open-ended and imaginative nature of the prompt may have encouraged aspirational or normatively desirable responses. The results should therefore be interpreted as articulated expectations rather than stable beliefs or actual classroom practices.

Third, the dataset is multilingual. Responses were translated into English for analysis, and although translations were checked for consistency, no full back-translation was conducted. Some nuances in meaning or cultural framing may therefore have been lost.

Fourth, the analytical approach involves partially dependent observations, as responses could be assigned to multiple categories. The statistical results should thus be understood as indicative patterns rather than definitive inferential tests. In addition, AI-assisted coding may introduce bias related to prompt design and model behaviour, despite human validation.

Despite these limitations, the study provides large-scale insights across multiple contexts into how students articulate expectations towards AI in education. Future research could address these limitations by including demographic data, longitudinal designs, and complementary methods.

4.6 Conclusion

Our findings challenge the assumption that AI inevitably undermines learner autonomy. Instead, students articulate clear expectations for support that enable thinking rather than replace it, and they frame AI through pedagogical roles rather than technical features.

In consequence, AI use in schools is not primarily a technical question, but a didactic one. What matters is not whether AI is present, but how its role is defined, framed, and regulated in relation to educational goals. In this respect, AI does not diminish the importance of teachers. On the contrary, it increases the demand for didactic clarity, professional reflection, and pedagogical leadership.

Beyond its substantive findings, the study also points to a practical implication for classroom practice. AI-supported analysis of open student responses can help make learners' conceptions visible on a scale and support responsive teaching. Used in this way, AI becomes not a shortcut around learning, but a tool for understanding it. Future research should examine how these expectation profiles and role conceptions develop over time and across subjects, and how an integration of AI as a pedagogically grounded resource can be further supported.

References

• 1

  AllenL. K.KendeouP. (2024). ED-AI Lit: an interdisciplinary framework for AI literacy in education. Policy Insights Behav. Brain Sci.11 (1), 3–10. 10.1177/23727322231220339

  • CrossRef
  • Google Scholar
  • View reference in article

• 2

  BaylorA. L.KimY. (2005). Simulating instructional roles through pedagogical agents. Int. J. Artif. Intell. Educ.15 (2), 95–115. 10.3233/IRG-2005-15(2)02

  • CrossRef
  • Google Scholar
  • View reference in article

• 3

  BernackiM. L.GreeneM. J.LobczowskiN. G. (2021). A systematic review of research on personalized learning: personalized by whom, to what, how, and for what purpose(s)?Educ. Psychol. Rev.33 (4), 1675–1715. 10.1007/s10648-021-09615-8

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  BroadfootP.RockeyJ. (2025). Generative AI and the social functions of educational assessment. Oxf. Rev. Educ.51 (2), 281–300. 10.1080/03054985.2025.2455549

  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  CarleyK.PalmquistM. (1992). Extracting, representing, and analyzing mental models. Soc. Forces70 (3), 601–636. 10.1093/sf/70.3.601

  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  ChanC.K.Y.HuW. (2023). Students' voices on generative AI: perceptions, benefits, and challenges in higher education. Int. J. Educ. Technol. High. Educ. 20, 43. 10.1186/s41239-023-00411-8

  • CrossRef
  • Google Scholar
  • View reference in article

• 7

  ChouC-YChanT-WChenZ-HLiaoC-YShihJ-LWuY-Tet al (2024). Defining AI companions: a research agenda—from artificial companions for learning to general artificial companions for global harwell. Res. Pract. Technol. Enhanc. Learn.20, 32. 10.58459/rptel.2025.20032

  • CrossRef
  • Google Scholar
  • View reference in article

• 8

  ConradtyC.BognerF. X. (2018). From STEM to STEAM: how to monitor creativity. Creat. Res. J.30 (3), 233–240. 10.1080/10400419.2018.1488195

  • CrossRef
  • Google Scholar
  • View reference in article

• 9

  ConradtyC.BognerF. X. (2019). From STEM to STEAM: cracking the code? How creativity & motivation interacts with inquiry-based learning. Creat. Res. J.31 (3), 284–295. 10.1080/10400419.2019.1641678

  • CrossRef
  • Google Scholar
  • View reference in article

• 10

  ConradtyC.BognerF. X. (2020). Conceptual change when growing up: frameset for role models?Int. J. Adolesc. Youth.25 (1), 292–304. 10.1080/02673843.2019.1622581

  • CrossRef
  • Google Scholar
  • View reference in article

• 11

  ConradtyC.BognerF. X. (2024). “Learning how to flow: paving the way for lifelong learning,” in Education and Human Development. Lifelong Learning—education for the Future World, eds. MeltzoffK.Gomez PalomaF. (IntechOpen), Vol. 19. 10.5772/intechopen.114171

  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  DangolA.WolfeR.ZhaoR.KimJ.RamananT.DavisK.KientzJ. A. (2025). Children’s mental models of AI reasoning: implications for AI literacy education. In Proceedings of the 24th Interaction Design and Children (pp. 106–123). ACM. 10.1145/3713043.3728856

  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  de PaoliS. (2024). Performing an inductive thematic analysis of semi-structured interviews with a large language model: an exploration and provocation on the limits of the approach. Soc. Sci. Comput. Rev.42 (4), 997–1019. 10.1177/08944393231220483

  • CrossRef
  • Google Scholar
  • View reference in article

• 14

  DuitR.TreagustD. F. (2003). Conceptual change: a powerful framework for improving science teaching and learning. Int. J. Sci. Educ.25 (6), 671–688. 10.1080/09500690305016

  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  ErduranS.LevriniO. (2024). The impact of artificial intelligence on scientific practices: an emergent area of research for science education. Int. J. Sci. Educ.46 (18), 1982–1989. 10.1080/09500693.2024.2306604

  • CrossRef
  • Google Scholar
  • View reference in article

• 16

  FuY.WengZ. (2024). Navigating the ethical terrain of AI in education: a systematic review on framing responsible human-centered AI practices. Comput. Educ. Artif. Intell.7, 100306. 10.1016/j.caeai.2024.100306

  • CrossRef
  • Google Scholar
  • View reference in article

• 17

  GiannakosM.AzevedoR.BrusilovskyP.CukurovaM.DimitriadisY.Hernandez-LeoD.et al (2025). The promise and challenges of generative AI in education. Behav. Inf. Technol.44 (11), 2518–2544. 10.1080/0144929X.2024.2394886

  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  GligoreaI.CiocaM.OanceaR.GorskiA-TGorskiH.TudoracheP. (2023). Adaptive learning using artificial intelligence in e-learning: a literature review. Educ. Sci.13 (12), 1216. 10.3390/educsci13121216

  • CrossRef
  • Google Scholar
  • View reference in article

• 19

  HanY.HongS.LiZ.LimC. (2025). Defining and classifying the roles of intelligent learning companion systems: a scoping review of the literature. TechTrends69 (3), 567–581. 10.1007/s11528-025-01058-0

  • CrossRef
  • Google Scholar
  • View reference in article

• 20

  HayesA. S. (2025). “Conversing” with qualitative data: enhancing qualitative research through large language models (LLMs). Int. J. Qual. Methods.24, 16094069251322346. 10.1177/16094069251322346

  • CrossRef
  • Google Scholar
  • View reference in article

• 21

  HolsteinK.McLarenB. M.AlevenV. (2019). Co-Designing a real-time classroom orchestration tool to support teacher–AI complementarity. J. Learn. Anal.6 (2), 27–52. 10.18608/jla.2019.62.3

  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  JohnsonW. L.RickelJ. W.LesterJ. C. (2000). Animated pedagogical agents: face-to-face interaction in interactive learning environments. Int. J. Artif. Intell. Educ.11, 47–78. Available online at: https://www.researchgate.net/publication/200772440_Animated_Pedagogical_Agents_Face-to-Face_Interaction_in_Interactive_Learning_Environments

  • Google Scholar
  • View reference in article

• 23

  KasneciE.SesslerK.KüchemannS.BannertM.DementievaD.FischerF.et al (2023). ChatGPT for good? On opportunities and challenges of large language models for education. Learn. Individ. Differ.103, 102274. 10.1016/j.lindif.2023.102274

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  KhlaifZ. N.AlkoukW. A.SalamaN.Abu EidehB. (2025). Redesigning assessments for AI-enhanced learning: a framework for educators in the generative AI era. Educ. Sci.15 (2), 174. 10.3390/educsci15020174

  • CrossRef
  • Google Scholar
  • View reference in article

• 25

  LanM.ZhouX. (2025). A qualitative systematic review on AI empowered self-regulated learning in higher education. NPJ Sci. Learn.10 (1), 21. 10.1038/s41539-025-00319-0

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 26

  MahieuB.SchlichP.VisalliM.CardotH. (2021). A multiple-response chi-square framework for the analysis of free-comment and check-all-that-apply data. Food. Qual. Prefer.93, 104256. 10.1016/j.foodqual.2021.104256

  • CrossRef
  • Google Scholar
  • View reference in article

• 27

  MaplesB.CeritM.VishwanathA.PeaR. (2025). How AI Companions Shape Learners’ Socio-Emotional Learning and Metacognitive Development. AI & Society. 10.1007/s00146-025-02737-5

  • CrossRef
  • Google Scholar
  • View reference in article

• 28

  MiaoF.HolmesW. (2023). Guidance for Generative AI in Education and Research. UNESCO. 10.54675/EWZM9535

  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  MillsN.HokH.DressenA.VeillasQ. (2025). The design and evaluation of an interactive AI companion for foreign language writing. Foreign Lang. Ann.58 (1), 40–69. 10.1111/flan.12790

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  NystrandM.WuL. L.GamoranA.ZeiserS.LongD. A. (2003). Questions in time: investigating the structure and dynamics of unfolding classroom discourse. Discourse Process.35 (2), 135–198. 10.1207/S15326950DP3502_3

  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  PacaciC.UstunU.OzdemirO. F. (2024). Effectiveness of conceptual change strategies in science education: a meta-analysis. J. Res. Sci. Teach.61 (6), 1263–1325. 10.1002/tea.21887

  • CrossRef
  • Google Scholar
  • View reference in article

• 32

  SajjaR.SermetY.CikmazM.CwiertnyD.DemirI. (2024). Artificial intelligence-enabled intelligent assistant for personalized and adaptive learning in higher education. Information15 (10), 596. 10.3390/info15100596

  • CrossRef
  • Google Scholar
  • View reference in article

• 33

  SampsonD. G.ZervasP.SotiriouS. (2011). Science education resources supported with educational metadata: the case of the OpenScienceResources web repository. Adv. Sci. Lett.4 (11), 3353–3361. 10.1166/asl.2011.2045

  • CrossRef
  • Google Scholar
  • View reference in article

• 34

  SchiffD. (2022). Education for AI, not AI for education: the role of education and ethics in national AI policy strategies. Int. J. Artif. Intell. Educ.32 (3), 527–563. 10.1007/s40593-021-00270-2

  • CrossRef
  • Google Scholar
  • View reference in article

• 35

  SchiltzH. K.ClarkeE.RosenN.De La RosaS. G.MasjediN.ChristopherK.et al (2024). A longitudinal mixed-methods characterization of family support from adolescence to young adulthood in autism and other developmental disabilities. J. Autism Dev. Disord.54 (9), 3225–3241. 10.1007/s10803-023-06070-y

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 36

  SotiriouS. A.BognerF. X. (2023). “Discovery space: a technology-enhanced classroom for deeper learning in STEM,” in Reimagining Education—the Role of E-Learning, Creativity, and Technology in the Post-Pandemic Era, ed. MistrettaS. (IntechOpen). 10.5772/intechopen.1002649

  • CrossRef
  • Google Scholar
  • View reference in article

• 37

  SotiriouS. A.LazoudisA.BognerF. X. (2020). Inquiry-based learning and E-learning: how to serve high and low achievers. Smart Learn. Environ.7, 29. 10.1186/s40561-020-00130-x

  • CrossRef
  • Google Scholar
  • View reference in article

• 38

  VeugenM. J.GulikersJ.den BrokP. (2021). We agree on what we see: teacher and student perceptions of formative assessment practice. Stud. Educ. Eval.70, 101027. 10.1016/j.stueduc.2021.101027

  • CrossRef
  • Google Scholar
  • View reference in article

• 39

  WatersA. R.JonesS. R.UppalapatiM.GududuruA.BonoM. H.HechtH. K.et al (2024). A content analysis of cancer-related changes in perceptions of self, relationships, and health among LGBTQI+ cancer survivors across the life course: findings from OUT: the national cancer survey. Psychooncology33 (12), e70044. 10.1002/pon.70044

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 40

  WilliamsR. T. (2024). The ethical implications of using generative chatbots in higher education. Front. Educ.8, 1331607. 10.3389/feduc.2023.1331607

  • CrossRef
  • Google Scholar
  • View reference in article

• 41

  ZhaiC.WibowoS.LiL. D. (2024). The effects of over-reliance on AI dialogue systems on students’ cognitive abilities: a systematic review. Smart Learn. Environ.11, 28. 10.1186/s40561-024-00316-7

  • CrossRef
  • Google Scholar
  • View reference in article