Most existing studies on AI in writing education focus on drafting, feedback, or assessment. Systematic reviews show that research on educational chatbots and generative AI has largely examined their role in providing feedback, personalizing instruction, or improving assessment quality, rather than supporting the early stages of idea generation (Debets et al., 2025; Feng et al., 2025; Lo et al., 2024; Marzano, 2025; Wu et al., 2025). Although this work has advanced understanding of how AI influences written products, it offers limited insight into how pupils engage cognitively and socially while developing ideas.

A smaller but growing strand of research examines Socratic or question-only chatbots. These systems are designed to support reasoning and idea development through dialogue, rather than text generation. Experimental studies show that these designs can stimulate engagement and reflection but also provoke resistance. Favero et al. (2024) demonstrate that a Socratic chatbot can foster critical thinking when pupils respond to questions and elaborate on their own ideas. In contrast, Blasco and Charisi (2024) find that although a Socratic GPT-4 tutor led to more frequent interaction among secondary-school pupils, it was often perceived as less helpful than a direct-help GPT-4 tutor and produced no measurable learning gains. Many pupils expressed frustration at its insistence on questioning rather than providing answers. Together, these findings highlight two aspects of Socratic AI: It can activate reasoning but also expose motivational and affective limits. At the university level, Umarova et al. (2025), similarly, find that learning outcomes depended less on a chatbot’s design than on how actively writers engaged with it. Related research on dialogic AI support more broadly suggests that reflective questioning can enhance metacognitive engagement even when the system is not strictly Socratic (Lai et al., 2025).

Across this literature, questioning is described as a powerful form of instructional support, but also as a potential source of interference. In a recent Norwegian classroom study, Strømman and Knutsen (2025) examined how fifth-grade pupils used Sokrates, a chatbot designed to foster reflection through guided questioning during story writing. For some pupils, the questions supported elaboration and richer description. For others, however, the chatbot began to steer the narrative by introducing characters and plotlines the pupils had not intended.

These findings point to a central design challenge in AI-supported prewriting: when a tool supports idea generation through dialogue, it must remain responsive to pupils’ developing intentions rather than shaping the content of the text. This tension is explicit in Strømman and Knutsen’s (2025) material, where one pupil responds to the chatbot’s intervention by insisting, “It is my story, not yours!”

The question of control in writing support is not new. Long before AI entered classrooms, composition researchers warned that overly directive feedback could erode pupils’ sense of authorship. Straub (1997) found that pupils valued comments framed as guidance or open questions but resisted those that reshaped their ideas. His findings suggest that feedback is most effective when it invites, rather than directs, revision; a principle that becomes even more critical when conversational agents are introduced into writing. The balance between guiding and taking over the task is particularly fragile before writing begins, when pupils are still forming ideas and often seeking direction.

Despite these insights, little is known about how pupils actually use question-based AI during prewriting or how such dialogues shape agency, elaboration and control before any text is written. Existing studies examine outcomes, user perceptions or drafting support, but they seldom analyze the turn-by-turn development of ideas in the earliest phase of writing. Much of the existing literature links AI support to improvements in written products, yet the prewriting phase itself has rarely been studied as an analytic object in its own right. In the present study, we therefore focus on the dialogue rather than the pupil’s final texts, treating prewriting as a distinct phase in which idea formation and agency can be observed directly. To address this gap, we examine how lower-secondary pupils interact with a question-only chatbot specifically designed to support prewriting by asking one question at a time and refusing to generate text.

Writing is understood in this study as both a cognitive process of planning and a socially mediated action. To analyze how pupils developed ideas in dialogue with the chatbot, we draw on two complementary perspectives: a cognitive view of writing as goal-directed problem solving (Flower and Hayes, 1981) and a sociocultural view of writing as mediated action (Wertsch, 1998). Together, these perspectives illuminate both how ideas are generated and organized and how agency and control are negotiated when a tool becomes part of the activity.

It has become increasingly common to integrate cognitive and sociocultural perspectives in studies of reading and writing (Klein and Boscolo, 2016; Uccelli et al., 2020). These approaches no longer view writing as either an internal mental process or a social practice but, rather, as a socio-cognitive activity in which individual planning and socially mediated processes intertwine. Following this broader trend, in the present study, we draw on Flower and Hayes (1981) and Wertsch (1998) to analyze how planning and mediation unfold in pupil–chatbot dialogues.

From a cognitive perspective, writing is a form of goal-directed problem solving (Flower and Hayes, 1981). During planning, writers generate ideas, retrieve knowledge, and set goals using cues from the task or their own memory to focus their attention and activate their relevant knowledge. These cues bridge intuitive thought and explicit language (Flower and Hayes, 1984), transforming loosely formed mental images into verbal plans. When a chatbot uses questions-only, they may function as such heuristic cues: They draw attention to sensory and memory-based detail, externalizing part of the cognitive work of planning while leaving wording and authorship to the pupil. In this sense, the chatbot can extend the planning space envisioned by Flower and Hayes while maintaining the boundary around what counts as the pupil’s own text (Hayes, 2012; Graham et al., 2024).

A sociocultural perspective adds another layer by also considering how people use cultural tools. Writing, as Wertsch (1998) argues, is always a mediated action: Writers act through cultural tools, and these tools organize the structure and dynamics of action by defining what kinds of actions are possible and how control is distributed. The concept of appropriation captures this tension between using a tool and having one’s actions organized through it. This concept concerns agency and how intentions and voices are negotiated when people rely on culturally and technologically pre-shaped means. In this study, such negotiation becomes visible when pupils encounter a system that refuses to do the writing for them. One pupil captured this struggle in a moment of frustration: “Can you just, for fuck’s sake, write the text for me?” The chatbot’s refusal returns agency to the pupil, forcing them to take ownership of the planning process. The chatbot’s design inevitably frames what counts as relevant detail, but the pupils still negotiate how to interpret, extend, or resist its cues. In the analysis, appropriation is therefore examined in terms of how pupils respond to, adapt, or resist the chatbot’s questions.

By bringing these perspectives together, we view prewriting as both cognitive planning and mediated negotiation. The chatbot’s fixed instructions—asking one question at a time, probing for detail, and refusing to generate text—create a bounded yet dialogic planning space. Within that space, pupils must decide how to take up, reshape, or reject its cues. Analyzing these decisions allows us to trace how idea generation unfolds under constraint and contestation. In this article, we therefore focus on the dialogues themselves, rather than the pupils’ final texts. Our aim is to capture how ideas took shape and evolved during prewriting: how they surfaced, stalled, or shifted through dialogue with the chatbot.

We ask the following research questions:

This is an exploratory qualitative classroom study using reflexive content analysis (Nicmanis, 2024) of pupil-chatbot dialogues. The study employs a single case design with 17 participants engaging with identical task conditions. Rather than seeking statistical generalization, the study aims for analytical generalization (Yin, 2018): detailed examination of how question-driven dialogue shapes prewriting processes under constraint. Norway provides a relevant context for examining AI-supported writing processes, as the education system has invested heavily in educational technology and provides one-to-one digital access for most pupils, supported by long-standing policy ambitions positioning Norway among the leading nations in digital education (Erstad, 2006; Klausen, 2020).

Writing holds a similarly central position as one of five basic skills in the Norwegian national curriculum. Process-oriented writing instruction has long been emphasized, with classroom research and assessment practices foregrounding planning, elaboration, and revision (Ongstad, 2002; Blikstad-Balas et al., 2018). Recent research shows that the national written examination in Norwegian—the only mandatory extended writing test—strongly shapes classroom writing practices (Eriksen et al., 2024). This dual emphasis on writing as process and digital competence creates a distinctive context for investigating AI-supported prewriting dialogue.

One lower-secondary school in a rural Norwegian municipality was recruited through convenience sampling (Bryman, 2016). The school principal forwarded the recruitment request to teachers, and one 8th-grade Norwegian Language Arts (L1) teacher volunteered to participate with their class. There was no prior relationship between the research team and the school, teacher, or pupils.

Seventeen pupils (aged 13–14; 12 girls and 5 boys) participated in the study. One additional male pupil was excluded by the teacher due to recent arrival and individualized learning needs. The class received 4 hours of Norwegian Language Arts instruction weekly and had, during the autumn term of 8th grade, worked extensively on narrative writing, engaging in activities focused on plot development, character description, and setting. According to the teacher, this was a typical 8th-grade class with the expected range of writing abilities. As pupils transition to lower-secondary school in 8th grade and formal grading only begins at this level, there were no prior achievement records documenting individual writing proficiency from primary school.

Informed consent was obtained from school administration, the teacher, all pupils, and their guardians in accordance with guidelines from Norwegian National Research Ethics Committees (2016). Participation was voluntary, no incentives were provided, and all data were anonymized.

No baseline writing assessments or additional demographic variables were collected. At the lower-secondary level in Norway, there are no standardized writing tests, and schools do not maintain systematic records of pupils’ writing proficiency or socioeconomic background. Writing instruction is integrated across the L1 curriculum rather than treated as a separate discipline with distinct grading practices (Norwegian Directorate for Education and Training, 2020). Because the analysis focused on interactional processes during prewriting dialogue rather than outcomes or individual characteristics, demographic information beyond age and gender was not collected.

A brief pilot phase conducted in January 2025 familiarized pupils with the chatbot’s question-only format prior to the main data collection.

The main data collection session took place in February 2025 during a regular L1 lesson. The writing task—“Describe your hometown to someone who has never been there before”—was introduced orally by the teacher and the first author. Pupils received an identical written start instruction via Microsoft Teams, which they pasted into the chatbot interface. The instructional sequence and data collection procedure are summarized in Figure 1.

Pupils engaged in a 45-min prewriting dialogue with the chatbot, followed by independent drafting without chatbot access. Pupils accessed the chatbot through chat.sikt.no, a Norwegian platform running GPT-4o via Microsoft Azure (February 2025 version). At the end of the session, pupils copied their complete chat logs into Word documents, which were collected immediately via USB drive. All original text was retained verbatim, including spelling errors, emojis, slang, and profanity.

The writing task was adapted from the Norwegian national 8th-grade writing assessment developed by the Norwegian Writing Centre. The task emphasized elaboration, relevant examples, and consideration of the reader’s informational needs. In the official scoring guide, high-level performance is defined by the expansion of ideas through explanation and detail rather than by listing features.

The teacher created a dedicated chatbot (“Friendly Wilma”) with a minimal system instruction (“you are a friendly and open chatbot”). Pupils pasted a start instruction into the chat interface that overrode this system prompt. The start instruction specified that the chatbot should ask one question at a time, encourage elaboration, and never produce text on behalf of the pupil (see Supplementary Appendix 1).

The chatbot generated contextually responsive questions (e.g., sensory, memory-based, or reflective prompts) depending on pupil input. It was explicitly instructed to refuse all requests to generate text, ensuring that all wording and formulations remained the pupil’s own. No additional materials were available during the chat.

During pilot testing, we observed that when the constraint was implemented only through a system prompt, the chatbot occasionally violated the rule under sustained pupil requests. When the same instruction was introduced as a visible start prompt, the constraint held consistently across interactions. For this reason, the start prompt was treated as a core design element of the intervention rather than a purely technical setting. The start prompt instructed the chatbot to ask one question at a time, encourage elaboration, and direct pupils’ attention toward sensory detail, memories, and personal significance (see Supplementary Appendix 1).

The chatbot generated questions responsively, adapting to each pupil’s preceding answer. This was achieved through the combination of the language model’s conversational capabilities and the explicit pedagogical constraints in the start prompt, which instructed the chatbot to guide pupils through sensory description, memories, and comparisons (Supplementary Appendix 1). The chatbot could therefore ask about memories following sensory descriptions and request elaboration when responses were brief, without requiring predefined question sequences.

The object of analysis was the prewriting dialogue. In writing research, prewriting is widely understood as a distinct phase in which writers explore, test, and refine emerging ideas prior to text production. Because the purpose of the study was to examine how ideas were formed, expanded, redirected, or resisted in interaction with a question-only chatbot, only the prewriting dialogues were included in the dataset.

The analysis followed a sequential, turn-by-turn approach informed by reflexive content analysis (RCA) (Nicmanis, 2024). Each dialogue was examined in its entirety, with attention to how pupils responded to the chatbot’s questions across successive turns and how interactional patterns developed over time. Interpretations were grounded in observable responses to the chatbot’s actual wording, with analytic claims warranted by what pupils did next in the dialogue.

The unit of analysis was the question–response (Q → R) pair, consisting of one chatbot question and the pupil’s immediate response. Both turns were included in the analysis. Coding proceeded in two layers, each addressing a distinct analytic focus.

Quality was pursued through transparency, credibility, and reflexivity. Transparency is supported by the explicit coding framework (Figure 2), the fixed start instruction used across all interactions (Supplementary Appendix 1), and detailed operational definitions with illustrative excerpts provided in Supplementary Tables S1, S2.

Credibility rests on grounding analytic claims in observable, turn-by-turn interaction. Interpretations were warranted by pupils’ subsequent responses in the dialogue, allowing analytic claims to be traced directly to interactional evidence rather than inferred intentions or assumed learning outcomes.

Coding was conducted by the first author using a reflexive, interpretive approach rather than a reliability-oriented one. Rigor was pursued through ongoing analytic strategies rather than post hoc checks (Morse, 2015). These strategies included freezing the start instruction prior to analysis, coding surface characteristics before functional interpretation, and systematically identifying and documenting deviant and negative cases, including instances where questions failed to generate elaboration or where pupils resisted the chatbot’s framing.

To strengthen analytic credibility, the complete coding scheme and selected transcripts were reviewed by the co-authors, who provided critical feedback on code definitions, category boundaries, and analytic interpretations. This process focused on conceptual clarity and categorical coherence rather than inter-rater agreement. Responsibility for final coding decisions remained with the first author.

The first author served as both researcher and observer during the classroom intervention, which may have influenced pupil behavior. To reduce observer effects, the classroom teacher introduced the task and remained present throughout the session. To address interpretive bias, analytic claims were anchored in observable interaction rather than inferred intentions, negative cases were retained and integrated into the analysis, and alternative readings and analytic uncertainties were documented in an audit trail throughout the research process.

All data are stored securely on an encrypted, password-protected computer and a locked external drive, accessible only to the first author.

To answer RQ1, “In what ways does a question-only chatbot facilitate or constrain pupils’ prewriting, particularly in generating descriptive content?” we examined how pupils responded to six functional question types across 201 coded exchanges in 17 dialogues (Table 1).

Two tendencies stood out. Sensory and memory questions most often led to elaboration within the same thematic frame, whereas reflective ones produced more variable outcomes, sometimes expanding the dialogue and sometimes restricting it. However, the pattern was far from uniform: Some pupils engaged actively with nearly every cue; others replied briefly or resisted the frame.

Across all the dialogues, the pupils produced 201 coded responses: Descriptive (72) and reflective (37) were most common; sensory (19) and memory-based (19) occurred less often; 31 were non-uptake, typically short refusals, such as “no” or “do not know”; and 14 were minimal. Two pupils tested the chatbot’s limits by asking it to write (seven times in total). Kasper asked for a summary twice.

Engagement varied widely: Some pupils, such as Christine and Nora, sustained extended exchanges that were rich in description and reflection, while others, such as Tina and Une, replied briefly or evaluatively. Dialogue length ranged from over 554 to 50 words, showing how the same question-only design produced sharply different levels of elaboration.

Most dialogues began with an opening question, which was typically phrased as “What is special about your hometown?” This initial move established a consistent starting point for all interactions, inviting the pupils to draw on their own experiences. As the conversations unfolded, the chatbot occasionally employed meta moves (statements that clarified its role, offered an apology, or redefined the task) to manage non-uptake or refusals. Such moments made the pedagogical boundaries visible (“I cannot write the text for you, but I can help you develop your ideas”).

The following sections examine the remaining functional categories (Encourages detail, Idea Generation, and Supports reflection), showing how particular types of questions sustained or halted the momentum after the initial frame was established.

To address RQ2, we examined how pupils’ responses to a question-only chatbot unfolded across entire prewriting dialogues. Analysis focused on how pupils took up, negotiated, or resisted the chatbot’s questions over time. Across the dataset, these interactional response patterns clustered into three dominant forms: stable descriptive engagement, negotiated relevance, and stalled or fragmented engagement (see Figure 3). The following sections present analytically representative cases illustrating each pattern.

Recent studies of question-based AI in classrooms show varied effects, from productive elaboration to resistance and frustration (Blasco and Charisi, 2024; Strømman and Knutsen, 2025). Across the dialogues, one pattern stood out clearly: sensory and memory questions were the most reliable triggers of elaboration. When asked what she could see or hear while walking up a path, Oda replied, “In summer you hear birds singing and, in winter, the snow falling from the trees.” Turid described walking her dog: “I hear the wind hitting the leaves of the trees, I smell fresh air, and I notice children playing outside.” In such moments, questioning anchored attention in concrete experience and opened a shared space for further development.

From a cognitive perspective, these sensory and memory questions functioned as retrieval cues (Flower and Hayes, 1981). Importantly, the chatbot externalized the cue to retrieve, not the retrieval itself. Pupils had to perform the work of translation: from perception to language, from vague impressions to articulated detail. The system did not describe the trees, the wind, or the children. It asked pupils to notice and name what they experienced. In this sense, the design does not eliminate cognitive offloading but redirects it. The cognitive work remains with the pupil, while the prompt to engage in that work is externalized.

Reflective questions showed a more conditional pattern. They were productive mainly when grounded in prior concrete detail. In Ole’s dialogue, for instance, an early question about how the sea’s sounds and smells affected his experience elicited the response “Did nothing.” This may reflect resistance, disengagement, or a lack of available representation. Later in the same dialogue, after Ole had recalled specific memories of the beach, similar reflective prompts generated more elaboration. Rather than indicating a fixed limitation, this pattern suggests that reflective questioning is more likely to be taken up when abstraction builds on representation.

These findings point to the importance of sequencing in prewriting support. The chatbot functioned as an external monitor (Hayes, 2012), persistently prompting attention, maintaining task focus, and declining to write. This externalized a function that skilled writers often perform internally during planning: sustaining inquiry, testing ideas, and deciding what to develop further. For some pupils, this structure supported cumulative elaboration. For others, it felt constraining or irrelevant. Engagement did not follow automatically from the design itself.

These results show that boundary stability was a necessary condition for sustained prewriting dialogue, but not a sufficient one. The same fixed constraint supported extended elaboration in some cases and stalled or resisted interaction in others. Understanding how pupils negotiated this boundary—sometimes productively, sometimes not—requires moving from question types to how pupils’ responses accumulated into dominant interactional patterns, which the next section takes up.

From a sociocultural perspective, the refusal to generate text functioned as a pedagogical boundary. Pupils had to decide whether to take up the invitation, negotiate its relevance, or reject the task altogether. In Wertsch’s (1998) terms, the tool structured what actions were possible, while control over meaning remained contested.

In this study, agency is not treated as an individual trait or a measure of motivation. Rather, it is understood interactionally, as it becomes visible in how pupils take up, negotiate, redirect, or resist the chatbot’s questions across turns. Agency is therefore inferred from observable response moves in the dialogue, not from pupils’ attitudes or presumed internal states. Minimal or stalled responses are read as interactional outcomes that constrain further idea development, rather than as a lack of agency per se.

The interactional patterns identified in section 3.2 show that support from a question-only chatbot was never simply delivered or received. Agency instead took shape through pupils’ moment-by-moment responses to the chatbot’s questions. Cognitively, the dialogue supported idea generation by externalizing planning cues. Sociocultural, the same turns functioned as mediated actions in which relevance, control, and legitimacy were continuously negotiated between pupil and tool.

In this section, we use negotiation as an umbrella term to describe how pupils positioned themselves in relation to the chatbot’s support, drawing on sociocultural accounts of mediated action and agency (Wertsch, 1998). Negotiation took different forms: appropriation, when pupils took up the questions productively for their own purposes; resistance, when they challenged the framing, relevance, or demands of the questions; and withdrawal, when they disengaged from the dialogic format altogether. These moves shaped agency not as an individual trait, but as an interactional accomplishment unfolding across turns.

Central to this negotiation was what we refer to as the pedagogical boundary: the chatbot’s absolute refusal to generate text on pupils’ behalf. This design directly addresses concerns raised by Baron (2023) about authorship in AI-supported writing: by refusing to generate text, the system made authorship not only non-negotiable but unavoidable. Pupils could not offload composition to the tool; they had to claim or reject ownership explicitly through their responses.

The chatbot followed a fixed instruction: it asked questions, encouraged elaboration, and refused to write. Many pupils accepted this role and worked productively within it. Frida’s statement, “I do not want to write about Skognes. I want to write about Furuvik,” illustrates productive appropriation through negotiation. She did not reject the question-driven format but redefined what counted as relevant, after which sustained descriptive work followed. Similarly, Nora’s “I’m not a weather expert” functioned as a playful deflection that recalibrated the level of abstraction without undermining engagement. In these cases, resistance operated as a resource for stabilizing the dialogue by aligning it with pupils’ own sense of task relevance.

These patterns extend Wertsch’s (1998) distinction between mastery and appropriation. While Wertsch conceptualizes appropriation as making a cultural tool one’s own, our analysis shows that in dialogic AI interaction, appropriation remains provisional and ongoing. It must be continuously re-established across turns as pupils and chatbot renegotiate relevance, authority, and purpose. Appropriation here is therefore not an endpoint, but an interactional process. The limits of this process become especially visible in Anette’s dialogue. At several points, she explicitly rejected both the relevance of the chatbot’s questions and the legitimacy of its role, culminating in a withdrawal from the interactional contract. Here, resistance no longer functioned as negotiation but as a refusal of dialogic participation, leading to interactional collapse.

Kasper’s dialogue represents a related but distinct boundary case. Unlike Anette’s affectively charged rejection of the chatbot’s questioning role, Kasper’s repeated requests for examples, summaries, or finished text challenged not the chatbot’s legitimacy, but the value of prewriting itself. His redirects sought to bypass exploration in favor of immediate product. Although less confrontational in tone, these moves similarly eroded dialogic viability by shifting the interaction away from planning and toward outcome. In both cases, the chatbot’s consistent regulatory responses made the pedagogical boundary visible through conflict rather than explanation.

These contrasts reveal that resistance does not function uniformly. In some dialogues, it stabilized interaction by recalibrating focus or demand; in others, it destabilized the task frame until descriptive work collapsed. We therefore describe the interactional patterns in Section 3.2 as fragile achievements: they required sustained mutual alignment across turns. A single abstract question arriving too early, repeated minimal uptake, or a moment of pupil disengagement could destabilize the entire exchange.

The question-only design operationalizes Straub’s (1997) principle that effective support invites rather than directs. The chatbot questioned but refused to reshape pupils’ ideas or take over their writing. By making this refusal absolute, the design transformed a pedagogical principle into a technical constraint—one that pupils could negotiate within but never circumvent.

These interactional patterns show that pupils’ engagement with a question-only chatbot cannot be reduced to acceptance or uptake. From a dialogic perspective, engagement involves entering and sustaining a shared space for thinking, where relevance and direction remain open rather than fixed (Wegerif and Casebourne, 2025). What becomes analytically salient, then, is whether pupils continue to treat the dialogue as a space worth inhabiting.

Within such a dialogic space, moments of irritation, refusal, or challenge do not necessarily signal immediate disengagement. Anette’s dialogue, however, illustrates a clear boundary case. While she initially participates, she ultimately rejects both the task and the form of support. This withdrawal does not represent an alternative form of engagement, but an interactional collapse of the dialogic arrangement. In this material, agency in AI-supported prewriting is negotiated through pupils’ interactional responses rather than through compliance with the tool, and resistance marks the point at which dialogic support can no longer be sustained.

A question-only chatbot can support prewriting only when its constraints are stable. In this study, stability depended on where the constraint was implemented: the refusal to generate text held consistently because it was introduced as part of the dialogue itself, rather than as a hidden system instruction. This did not produce compliance. Several pupils tested, challenged, or resisted the constraint. What it ensured was a predictable interactional boundary that pupils could respond to, negotiate with, or reject.

Three design principles proved consequential. First, constrained functionality: the chatbot consistently refused to generate text, keeping formulation with the pupil. Second, instructional placement: embedding the constraint in the dialogue ensured that it remained operative across turns. Third, question-driven support: rather than offering suggestions or examples, the chatbot directed attention through questions that invited pupils to notice, recall, and articulate their own ideas.

These principles may extend across subjects and genres, although this study examined only descriptive writing. In narrative tasks, questions might invite attention to character motivation or setting. In argumentative writing, they might prompt pupils to articulate claims, evidence, or counterarguments. In science tasks, questions could focus on observation, explanation, or causal reasoning. What changes is the focus of the questioning; what remains constant is the division of labor: pupils do the writing, and the chatbot organizes attention through questions.

Whether this approach works depends partly on the teacher’s ability to design prompts that reflect the specific demands of the task. A poorly designed prompt, one that asks generic questions or repeats itself without adapting, will produce generic responses or stalled dialogue. Well-designed start prompts anticipate the kind of thinking required and guide attention accordingly. Teachers should also frame the activity before pupils begin, making explicit what the tool can and cannot do, and why that boundary matters.

When dialogue stalls or pupils resist, this should not be read simply as disengagement. Redirecting focus, challenging relevance, or requesting summaries are ways pupils negotiate what counts as meaningful work. Using such moments as shared objects of discussion can help pupils see how agency and task framing are constructed through dialogue.

Question-driven support is not appropriate for all prewriting contexts. When pupils lack basic content knowledge, questioning alone will not generate ideas. When pupils already produce detailed descriptions but struggle with abstraction, sustained sensory questioning may reinforce rather than address the problem.

The study does not suggest that pupils internalize questioning strategies after a single encounter with a question-only chatbot. Rather, the findings show how structured questioning can make planning processes visible and available during prewriting. Over time, and with repeated use, such dialogic practices may support planning habits that pupils can later draw on when writing independently.

A longer-term possibility, beyond the scope of the present study, is that pupils begin to shape this kind of support themselves. A pupil who can formulate a prompt such as “Ask me about sensory details, but do not describe anything for me” is not showing internalization, but an emerging ability to recognize what kind of external support helps their thinking.

This study contributes an empirical account of prewriting as an interactional process, focusing on how ideas take shape, stall, or shift through dialogue before any text is produced. Rather than evaluating writing outcomes, the analysis treats prewriting dialogue as an object of inquiry in its own right. The findings should therefore be read as an analytic generalization about dialogic prewriting under constraint, not as evidence of instructional effectiveness or as a basis for causal claims about writing quality.

By isolating the planning phase, the analysis foregrounds how questioning, uptake, resistance, and withdrawal shape what becomes available for writing. Claims about learning outcomes or text quality would require different analytic criteria and theoretical framing and fall outside the scope of the present article. How interactional patterns during prewriting relate to subsequent textual production remains an important question for future research.

The findings also point to several open empirical questions. Question-driven dialogue supported sustained elaboration for some pupils but not for others, and this variation warrants closer examination. Differences may relate to prior writing experience, how pupils interpret dialogic work, or contextual factors shaping their willingness to remain in exploratory talk. The analysis also raises questions about transfer. It remains unclear whether pupils who engage productively with external questioning later draw on similar strategies when writing independently, for example in exam situations where AI tools are unavailable.

In addition, the findings suggest that in descriptive writing, reflective questions were more likely to be taken up when grounded in concrete representation. Whether similar sequencing applies in genres where abstraction may precede detail, such as argumentative or analytical writing, remains an open empirical question.

Several methodological and contextual constraints delimit the scope of the findings. All pupils worked with an identical start prompt, which limits insight into how variation in prompt design might shape interactional trajectories. In some dialogues, persistent encouragement to elaborate after brief responses contributed to repeated follow-up questions without a clear exit condition, occasionally resulting in stalled interaction. These patterns point to the need for further research on how persistence and restraint are balanced in question-driven support.

Finally, the study was conducted in an educational context where process-oriented writing pedagogy and high digital access shape expectations about prewriting and exploration. Transferability to test-driven systems or to contexts with limited technical infrastructure therefore cannot be assumed. The analysis captures a snapshot of initial use within a single session and does not address how repeated use over time might alter pupil strategies, teacher framing, or the perceived value of dialogic prewriting.