Paper deep dive
From Intention to Text: AI-Supported Goal Setting in Academic Writing
Yueling Fan, Richard Lee Davis, Olga Viberg
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 96%
Last extracted: 4/27/2026, 6:52:04 PM
Summary
This paper introduces WriteFlow, an AI-powered, voice-based writing assistant designed to support reflective academic writing through goal-oriented interaction. Developed as a Google Docs add-on, WriteFlow uses a Wizard-of-Oz approach to facilitate metacognitive regulation, allowing users to articulate, refine, and monitor writing goals through dialogic interaction. A formative study with 17 participants identified key challenges in goal management and authorship, leading to five design requirements. An expert user evaluation (n=12) demonstrated that WriteFlow effectively scaffolds reflection-in-action, helps maintain goal-text alignment, and supports iterative goal refinement, thereby enhancing the writer's agency and metacognitive awareness during the academic writing process.
Entities (11)
Relation Signals (8)
Yueling Fan â affiliatedwith â KTH Royal Institute of Technology
confidence 100% · Yueling Fan 1 ... KTH Royal Institute of Technology
WriteFlow â facilitates â Goal Setting
confidence 100% · WriteFlow frames AI interaction as a dialogic space for ongoing goal articulation, monitoring, and negotiation
WriteFlow â isa â AI voice-based writing assistant
confidence 100% · This study presents WriteFlow, an AI voice-based writing assistant designed to support reflective academic writing through goal-oriented interaction.
Yueling Fan â isaffiliatedwith â KTH Royal Institute of Technology
confidence 100% · Yueling Fan 1 ... Department of Media Technology and Interaction Design, KTH Royal Institute of Technology
WriteFlow â isbasedon â Google Docs
confidence 100% · We present WriteFlow, a Google Docsâbased writing assistant
WriteFlow â supports â Metacognition
confidence 90% · WriteFlow scaffolds metacognitive regulation and reflection-in-action by supporting iterative goal refinement
WriteFlow â supports â Metacognitive Regulation
confidence 90% · Findings from a Wizard-of-Oz study with 12 expert users show that WriteFlow scaffolds metacognitive regulation and reflection-in-action
WriteFlow â â
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:This study presents WriteFlow, an AI voice-based writing assistant designed to support reflective academic writing through goal-oriented interaction. Academic writing involves iterative reflection and evolving goal regulation, yet prior research and a formative study with 17 participants show that writers often struggle to articulate and manage changing goals. While commonly used AI writing tools emphasize efficiency, they offer limited support for metacognition and writer agency. WriteFlow frames AI interaction as a dialogic space for ongoing goal articulation, monitoring, and negotiation grounded in writers' intentions. Findings from a Wizard-of-Oz study with 12 expert users show that WriteFlow scaffolds metacognitive regulation and reflection-in-action by supporting iterative goal refinement, maintaining goal-text alignment during drafting, and prompting evaluation of goal fulfillment. We discuss design implications for AI writing systems that prioritize reflective dialogue, flexible goal structures, and multi-perspective feedback to support intentional and agentic writing.
Tags
Links
- Source: https://arxiv.org/abs/2604.15800v1
- Canonical: https://arxiv.org/abs/2604.15800v1
Trouble viewing inline? Open PDF directly â
Full Text
42,839 characters extracted from source content.
Expand or collapse full text
From Intention to Text: AI-Supported Goal Setting in Academic Writing Yueling Fan 1 , Richard Lee Davis 2,3 , and Olga Viberg 1,3(B) 1 Department of Media Technology and Interaction Design, KTH Royal Institute of Technology, Stockholm, Sweden yueling@kth.se 2 Department of Digital Learning, KTH Royal Institute of Technology, Stockholm, Sweden 3 Digital Futures, Stockholm, Sweden rldavis,oviberg@kth.se Abstract. This study presents WriteFlow, an AI voice-based writing assistant designed to support reflective academic writing through goal- oriented interaction. Academic writing involves iterative reflection and evolving goal regulation, yet prior research and a formative study with 17 participants show that writers often struggle to articulate and manage changing goals. While commonly used AI writing tools emphasize effi- ciency, they offer limited support for metacognition and writer agency. WriteFlow frames AI interaction as a dialogic space for ongoing goal ar- ticulation, monitoring, and negotiation grounded in writersâ intentions. Findings from a Wizard-of-Oz study with 12 expert users show that WriteFlow scaffolds metacognitive regulation and reflection-in-action by supporting iterative goal refinement, maintaining goalâtext alignment during drafting, and prompting evaluation of goal fulfillment. We dis- cuss design implications for AI writing systems that prioritize reflective dialogue, flexible goal structures, and multi-perspective feedback to sup- port intentional and agentic writing. Keywords: Academic writing· Goal setting· AI· Reflection-in-action · Metacognition. 1 Introduction Academic writing is a cornerstone of higher education, serving not only as a medium for assessment but as a powerful engine for learning, knowledge con- struction, and intellectual development. Rather than reproducing information, students engage in writing as a process of knowledge transformation, negotiating the dynamic interplay between rhetorical challenges (how to communicate ideas) and content-related challenges (what to communicate) [4]. Through this process, tacit, experiential, and fragmented knowledge can be externalized, refined, and made transferable across contexts [23]. Academic writing further fosters strategic and self-regulated cognitive skills [12], including goal setting, planning, monitor- ing, and revisingâcapabilities that are critical for scholarly inquiry, professional arXiv:2604.15800v1 [cs.HC] 17 Apr 2026 2Y. Fan et al. practice, and lifelong learning. However, despite its importance many students struggle to meet the demands of academic writing [17,20], highlighting a persis- tent gap between the recognized importance of writing and studentsâ ability to effectively engage in it. With the rise of large language models (LLMs), the ways students engage in academic writing practices have started to change. For example, systems such as ChatGPT show strong reasoning and open-ended text generation capabilities [26,2], and have become increasingly embedded in studentsâ learning [11]. Yet, growing evidence suggests that reliance on such tools may undermine learning by encouraging cognitive offloading and reducing metacognitive engagement [9]. In academic writing, these risks are particularly acute, since its value lies not only in text production but in sustained reflection, reasoning, and iterative goal revision [10,12]. Academic writing is a recursive and cognitively demanding process requiring writers to manage evolving goals, integrate cross-sectional ideas, and continu- ally evaluate and revise arguments [10,12]. However, most commercial LLM in- terfaces are optimized for linear, turn-based dialogue, which poorly aligns with non-linear writing processes. Revisiting earlier reasoning or managing multiple concurrent goals is cumbersome, often resulting in surface-level interactions that limit metacognitive regulation, which is a strong predictor of academic success [24]. These challenges can be seen through the lens of self-regulated learning (SRL), which emphasizes learnersâ active metacognitive regulation across fore- thought, performance, and reflection phases [31], with goal setting as a central mechanism [30]. In academic writing, goals are continuously revised as ideas evolve, making dynamic goal regulation essential for preserving the epistemic value of writing with AI. Recent AI-driven writing tools have begun incorporating SRL lens to sup- port self-reflection and critical evaluation [15,25], yet do not explicitly support iterative goal adjustment. This study aims to fill this gap by examining how AI- supported systems can scaffold reflection and goal setting in academic writing. We present WriteFlow, a Google Docsâbased writing assistant co-designed to support reflection-in-action [21] through AI-mediated dialogue, structured goal generation, and goal-alignment tracking. Building on prior work showing con- versational interaction as a powerful design resource for fostering reflection [3], WriteFlow reconceptualizes chat-based interaction as a space for goal negotiation and metacognitive regulation. This study contributes to research on AI-supported academic writing by (1) providing a formative account of studentsâ writing goalârelated stress and cur- rent patterns of LLM use; (2) introducing WriteFlow, a voice-based AI writing assistant that scaffolds metacognition and self-regulation through iterative con- struction and monitoring of writing goals; (3) providing empirical evidence of how WriteFlow supports tracking alignment between stated goals and emerging text; and (4) deriving design implications for humanâAI writing systems that support evolving goal setting and metacognitive scaffolding. From Intention to Text: AI-Supported Goal Setting in Academic Writing3 2 Background 2.1 Goal Setting for Self-Regulated Academic Writing Academic writing is guided by hierarchically structured goal structures that in- clude abstract intentions (e.g., audience awareness) and concrete subgoals (e.g., revising a paragraph) [10,12]. Expert writers generate richer goal structures and flexibly revise them as task constraints evolve [12]. Research shows planning and quality-oriented goals improve text quality and promote higher-level revi- sion behavior [1]. Specific, proximal, and appropriately challenging goals enhance metacognition, motivation, and overall writing performance [22,19,5]. Process- oriented goals are particularly effective, compared to product-focused goals, process goals combined with feedback support learning of writing strategies, strengthen self-efficacy, and promote transfer [22]. Although automated writing evaluation systems provide personalized feedback [13], they offer limited sup- port for reflection on writersâ underlying intentions or monitoring evolving goal structures during the writing process. 2.2 AI Writing Tools and Metacognitive Support Recent AI writing tools have increasingly incorporated support for metacog- nitive processes during academic writing. Systems such as VISAR [29] enable writers to construct hierarchical goal structures during the planning phase, help- ing them articulate and organize content goals before drafting begins, though research indicates such interfaces can increase cognitive load [18]. Other tools fo- cus on supporting metacognitive reflection during revision and feedback stages. Friction [28] helps writers formulate actionable revision goals when editing exist- ing drafts, while ALure [16] scaffolds self-regulated learning through structured prompts that encourage writers to reflect on their strategies and progress. Re- verse outlining approaches [7] enable retrospective assessment of whether written text aligns with intended structure, promoting reflective revision. Research on feedback timing suggests that continuous, in-action feedback better supports learning than post-hoc evaluation alone [14,8], though such approaches also risk fostering dependency if not carefully designed. Despite these advances in metacognitive support in planning and revision stages, existing tools do not help writers track the evolving relationship between their stated goals and emerging text during the drafting process itself. Writers lack explicit mechanisms to notice when drift occurs, i.e., when the text being produced no longer serves the goals originally intended. This gap is particularly acute in AI-assisted writing contexts, where generated content may subtly pull writers away from their intentions without them recognizing the misalignment until substantial revision is required. Writers need support not only for setting goals (planning tools) and evaluating completed text (revision tools), but also for maintaining awareness of goal-text alignment throughout drafting, enabling them to decide whether to realign their text with original goals or intentionally revise those goals in light of emerging insights. 4Y. Fan et al. Addressing this gap, we present empirical findings on how an AI-based writ- ing assistant can be designed to support metacognitive scaffolding through iter- ative goal setting and monitoring. This study consists of (1) a formative study informing the co-design of WriteFlow and (2) an expert user evaluation. 3 Formative Study To understand how adult writers use AI when addressing academic writing chal- lenges, we conducted a survey with 17 students (8 female, 8 male, 1 non-binary; aged 22â34). The open-ended online survey 4 was administered between May 22 and 27, 2025 and included ratings of ten writing challenges grounded in Cognitive Process Theory of Writing which posits that writing is a non-linear, goal-directed, and recursive mental process rather than a strictly staged product [12], along with questions about coping strategies and AI use. Responses were analyzed using Reflexive Thematic Analysis [6]. The study identified three key writing challenges and 20 coping strategies (detailed descriptive statistics and thematic analysis results are available in our OSF repository 5 ). The rating data showed that the most stressful challenges were evolving writing goals and setting writing goals, reflecting difficulties in revising plans as new ideas and sources emerge. Participants reported using outlining, documentation, and AI tools to track evolving ideas, test structural changes, and maintain alignment with cen- tral arguments. These strategies supported reflection and reduced uncertainty during revision. In their open-ended responses, the study participants also high- lighted a third key challenge: preserving authorship. Some participants (n = 6) reported primarily using ChatGPT in a human-in-the-loop manner, leveraging it for ideation, tone refinement, and feedback interpretation. However, other participants (n = 9) expressed concerns about overreliance, authenticity, and ownership. Based on these findings, we derived five design requirements (R) for AI-supported reflective academic writing: (R1) facilitate goal articulation; (R2) support iterative goal refinement; (R3) enable organization and revisiting of ideas relative to goals; (R4) preserve writer voice and meaning; and (R5) ensure AI feedback is transparent, revisable, and aligned with user intent. To- gether, these requirements underscore the centrality of human judgment and agency in AI-assisted academic writing. 4 Survey instrument, participant demographics, and detailed findings for the formative study: OSF repository, https://osf.io/ba6d2/overview?view_only= e69b00a3acfd42529d3fb2c9c10b1ef7 5 Survey instrument, participant demographics, and detailed findings for the formative study: OSF repository, https://osf.io/ba6d2/overview?view_only= e69b00a3acfd42529d3fb2c9c10b1ef7 From Intention to Text: AI-Supported Goal Setting in Academic Writing5 Fig. 1. The overview of WriteFlow, a Google Docs add-on for goal-oriented academic writing. WriteFlow interface consists of a voice agent (A) and a sidebar panel with three pages: Writing Task (B), AI Chat (C), and My Goals (D). Users can upload Google Docs and communicate with the voice agent at any stage of writing to discuss their writing directions. The agent then generates writing goals to help them plan, track, and monitor their writing process. 6Y. Fan et al. 4 System Overview Figure 1 presents WriteFlowâs workflow and interface 6 . Users provide writing requirements and upload drafts, which the system uses to contextualize task understanding. Through Voice Mode, users discuss their writing plans with an AI-mediated conversational agent, which generates writing goals aligned with their intentions. Goals are stored on the My Goals page, where users can track progress, receive targeted suggestions, and review post-completion evaluations. WriteFlow also provides an Outline view that supports creating, revising, and comparing multiple outline versions across drafting stages. Goal Setting and Monitoring. WriteFlow supports planning and self- regulated writing by scaffolding goal articulation, refinement, and progress mon- itoring. During voice-based interaction, the system helps users externalize ideas and translates them into adaptive writing goals (R1, R3). Self-evaluation cards enable users to iteratively revise goals (R2), while progress tracking (R5) and suggestion cards support focused execution and sub-goal formation. Goal-Text Alignment Evaluation. Central to WriteFlowâs design is the Goal Completion Evaluation feature, which directly addresses the challenge of tracking alignment between evolving goals and emerging text. After goal com- pletion, the system evaluates alignment between goals, outlines, and written content to support reflective revision and preservation of authorial intent (R4). The Outline view further supports flexible goal evolution by allowing users to create and compare multiple outline versions across drafting stages, enabling tracking of how writing plans change over time. 5 User Evaluation We used WriteFlow as a design probe and conducted an exploratory Wizard- of-Oz study to answer the following research questions: RQ1. In what ways does the use of WriteFlow support usersâ metacognitive goal-oriented processes during academic writing, and what design refinements are needed, based on their feedback? RQ2. How does using WriteFlow influence writersâ sense of agency and their reliance on AI during academic writing? 5.1 Participants A total of 12 writers with expertise in humanâcomputer interaction (HCI) par- ticipated in the study (9 female, 3 male; ages 23â29). All had experience in English academic writing, and 7 had prior academic publication experience in the HCI field. Participants were intentionally recruited from the HCI domain, as they represent a population for whom reflective academic writing is a routine yet demanding practice and who possess the analytical skills required to crit- ically interrogate interactive system behavior. Their familiarity with academic 6 WriteFlowProtoPieprototype:https://cloud.protopie.io/p/ 3176a8c0ab9ad1f9e99b0910 From Intention to Text: AI-Supported Goal Setting in Academic Writing7 Fig. 2. Procedure of the user study. writing conventions and the design and evaluation of interactive systems enabled an informed assessment of WriteFlowâs support for goal articulation, authorial voice, and user control. Also, participants exhibited diverse AI use practices: five reported very frequent use of AI for writing, four reported frequent but more limited use, and three reported occasional use for specific stages of the writ- ing process. This variation allowed us to examine how prior experience with AI shaped engagement with and perceptions of the system during real-world writing tasks. 5.2 Study Design Participants interacted with WriteFlow, a high-fidelity ProtoPie prototype sup- porting multimodal input (text, document upload, and simulated voice). Al- though presented as an autonomous AI assistant, all responses were controlled in real time by a human facilitator (Wizard-of-Oz). Participants were asked to imagine they were completing a realistic academic task: writing a TikTok interaction critique from a user experience perspective. All study materials, in- cluding the task instructions, prompts described below, are available in an OSF repository 7 . WriteFlow Setup. We used GPT-4o to generate all AI responses. All prompts were pre-defined and pilot tested to ensure consistency and task relevance. To help participants quickly engage with the writing task, we pro- vided pre-prepared literature review notes on interaction criticism frameworks and TikTok case studies. Participants were encouraged to use these materials but could also incorporate their own ideas. We focused on how participants used WriteFlow during the writing process, particularly their strategy use and cog- nitive processes during the WriteFlow-assisted academic writing process, rather than evaluating final written outputs. 5.3 Procedure The study involved two writing tasks followed by post-study semi-structured interviews (Fig. 2). Each participant spent 120-150 minutes total and received a gift voucher as compensation for their time. All sessions were conducted remotely via Zoom and were both screen- and audio-recorded for accurate transcription. Before the study, participants provided demographic information and watched 7 https://osf.io/xjw6f/overview?view_only=f7779496c1ba4bbab7c928767a5cd7f2 8Y. Fan et al. a brief video demonstrating WriteFlowâs goal-setting workflow. The study dura- tion (over 2 hours) enabled deeper interaction with the system. At the beginning of the study, participants completed a written consent form and were introduced to the writing assignment. They were given 10 minutes to familiarize themselves with the provided literature review notes. Participants completed two writing tasks while thinking aloud. In Task 1 (Goal Generation Before Writing), they discussed their initial ideas with the AI, which generated writing goals. They evaluated and chose to accept or reject the suggested goals (minimum three goals). In Task 2 (Goal Revision), participants identified misalignments between a pre-written paragraph and a revised outline. A semi-structured interview fol- lowed to understand participantsâ perceptions and experiences with WriteFlow. 5.4 Data Analysis Recordings of usersâ (n = 12) interactions with WriteFlow (a total of 18 hours) and post-study semi-structured interviews (a total of 6 hours) were transcribed and analyzed using reflexive thematic analysis [6] by two researchers. We first familiarized ourselves with the transcripts by repeatedly reading and taking an- alytic notes on participantsâ transcripts. Then we open-coded how participants experienced WriteFlow during their writing tasks and how it succeeded or failed to support writersâ self-reflection. This coding was conducted in an inductive, data-driven manner, staying close to participantsâ language and concrete experi- ences. After the first few rounds of coding and discussion, the authors observed that most of the issues participants focused on when evaluating whether Write- Flow was helpful closely resonated with the five design requirements (R1âR5). For example, in their think-aloud protocols, participants frequently commented on how specific features supported their efforts to articulate and refine goals, organize and reuse ideas in the notes, preserve their authorial voice, etc. Build- ing on these observations, the first author then revisited the previously derived design requirements and we began to treat (R1âR5) as an analytic lens to at- tend more closely to how participantsâ experiences aligned with, extended, or challenged each requirement. We also extended our coding to capture additional design implications and considerations that participants proposed. Subsequently, the researchers conceptualized a set of themes and met regularly to check for disconfirming cases, merge or split themes when needed. Ultimately, we present two main themes in the Results section. 6 Results 6.1 Goal-Oriented Workflow as Metacognitive Scaffolding Facilitated Goal Articulation (R1). WriteFlow was experienced to support participants in clarifying "vague" or "underdeveloped" ideas (U1, U2, U4, U12) and generating "new perspectives" aligned with their writing intentions (U2, U4, U12). U12 for instance, appreciated voice input for enabling more fluid thinking, From Intention to Text: AI-Supported Goal Setting in Academic Writing9 facilitating her initial idea formation: âI might say a lot of nonsense at first, but I become clearer and clearer in the processâ. U11 further highlighted that "goal-setting helped define the scope of AI interaction within the broader writing task", making conversations more "focused" and "purposeful." Additionally, U3 and U8 suggested WriteFlow could be improved by asking follow-up questions or recommending relevant readings based on the ongoing conversation. Support Iterative Goal Refinement (R2). Several writers reported that WriteFlow enabled them to flexibly and continuously adjust their writing plans throughout the writing process (U1, U2, U5, U6, U7, U9, U11). U2 explained, âI can easily modify my writing plan at any stage. I donât need to be scared of losing the direction.â Similarly, U5 noted: âyou can go through and filter or refine those goals yourself,â highlighting the WriteFlowâs support for iterative goal refinement. U7 and U8 reported the Track Content feature improved writing efficiency by automatically linking goals to relevant text segments. This function was par- ticularly valuable during paper revision: âYouâl have many goals based on the reviewerâs feedback, each requiring changes in different partsâ (U12). U2 simi- larly stressed the feature helped her quickly navigate longer papers with multiple goals. Additionally, three other participants expressed a desire for the ability to directly navigate to source paragraphs within goal-linked documents (U2, U8, U9). Participants (n = 7) also emphasized the importance of visualizing the hierar- chical relationships between writing goals, suggesting that such representations could reduce cognitive overload and support the development of a more coher- ent and comprehensive goal network. To further improve goal management, U8 proposed that completing a parent goal should automatically mark its sub-goals as âcompletedâ, while others suggested categorizing goals by issue type (e.g., structure, citation) (U7) or by difficulty level (U11). U11 noted a preference for prioritizing more challenging goals first, indicating the value of flexible goal prioritization strategies. Enable organization and revisiting of ideas relative to goals (R3). Goal completion evaluation appeared to enhance metacognitive awareness by prompting writers to assess whether their drafts fulfilled their intended goals. Participants reported that WriteFlow helped them identify missing content and generate goal-aligned revisions (U1, U2, U3, U4). U1 remarked, âThe voice assis- tant tells you what youâre missing and what youâre not missing,â explaining that this feedback supported a more holistic view of her draft and made it easier to reflect on whether each section served its intended purpose. Similarly, U5 noted that although she might not typically initiate self-evaluation, WriteFlow âactively prompts you to do evaluations,â increasing her awareness of goal completion and progress monitoring. U4 noted the system helped her âgrasp the blueprint of the whole essayâ and become âmore conscious about whether what Iâm writing actually aligns with my goals, and what role it plays in the whole essay.â 10Y. Fan et al. Participants who previously described their writing as "divergent" or "un- structured" (U3, U11) expressed that WriteFlowâs goal-oriented workflow aligned more closely with the structured approaches of their co-writers or supervisors, which they aspired to adopt. Across participants, many emphasized that clari- fying goals and overall structure helped them maintain focus on their intended direction throughout the writing process (U2-U8, U11, U12). At the same time, some participants reported cognitive and motivational challenges associated with AI-generated evaluations. U1, U7, and U12 described feeling fatigued by lengthy descriptive feedback, while U12 questioned the AIâs ability to accurately evaluate her writing, noting that her personal standards often exceeded what AI could assess. U7 expressed a preference for constructive and supportive feedback over overly critical evaluations. Rather than receiving long textual critiques, U7 and U12 preferred the AI to provide high-quality worked examples that they could use for comparison and adaptation. Finally, U3 and U4 cautioned that single-perspective or overly complete responses could constrain divergent thinking and limit deeper exploration. To preserve writer agency and support reflective thinking, they suggested that WriteFlow should offer multiple perspectives or alternative options, encouraging users to critically evaluate suggestions rather than passively accept them. 6.2 Writer Agency and Critical Engagement with AI Feedback Preserve Writer Voice and Meaning (R4). Participants described Write- Flow as offering a greater sense of authorial control, referring to the extent to which users retain agency over the content and direction of their writing when in- teracting with AI. Several participants commented that WriteFlow helped them make more informed decisions by providing explanations behind each response (U4, U5). For example, as U5 highlighted: âThis tool allows decision-making at every decision point. When the AI provides something thatâs incorrect or off-track, it lets the user make a direct and convenient choice, like accept or reject. . . It gives you a comparison and a framework for evaluation. After evalu- ating, then I can decide whether to add that goal to the list.â Participants (U2, U3, U4) also shared that WriteFlow supported ongoing evaluation of whether their writing stayed aligned with their original intentions, as shared by U2 âI think I would prefer to use this system to help me to maintain my intention in the writingâ. U2 and U3 perceived WriteFlow as a supportive writing partner, one that encouraged them to propose confusion, disagreement, and alternative statements. This seemed to shift them from passively receiving AI-generated content (as with ChatGPT) to actively negotiating meaning and structure, thereby reinforcing their agency and voice throughout writing (U2, U3). Transparent Feedback to Support Critical Evaluation of AI Out- puts (R5). In this study, participants consistently evaluated and verified AI- generated outputs, and this verification process itself stimulated a deeper reflec- tion. This suggests that traceability and transparent reasoning are key not only for supporting informed decision-making, but also for fostering reflective writing From Intention to Text: AI-Supported Goal Setting in Academic Writing11 practices in human-AI interaction. Six participants reported that the systemâs use of concrete, sufficiently detailed evidence and clear, material-grounded rea- soning increased the perceived trustworthiness and reliability of its suggestions. U9 further proposed that visually representing the AIâs inference process could better support human verification, noting that tracing this reasoning might also help writers enter a more reflective cognitive state. U11 emphasized the value of explainable AI, expressing the expectation that the system would âexplain why this goal makes sense,â with reference to the source material. 7 Discussion This study aimed to design and evaluate WriteFlow, a human-informed, AI- driven writing assistant for reflective academic writing, with the aim of under- standing how goal-oriented interaction design can scaffold writersâ metacogni- tive regulation and foster agency. Using WriteFlow as a design probe in an ex- ploratory Wizard-of-Oz study, we addressed two research questions: (RQ1) how WriteFlow supports metacognitive, goal-oriented writing processes and what de- sign refinements are suggested by user feedback, and (RQ2) how interacting with WriteFlow influences writersâ sense of agency and reliance on AI during academic writing. Overall, WriteFlow was experienced not merely as a text-generation tool, but as a reflective partner that foregrounded academic writing as an intentional, goal-driven cognitive activity. This aligns with the Cognitive Process Theory of Writing, which conceptualizes writing as a recursive interaction between plan- ning, translating, and reviewing rather than a linear sequence of stages [10,12]. 7.1 Metacognitive Goal-Oriented Writing Process with AI Addressing RQ1, the findings show WriteFlow supports metacognitive, goal- oriented writing by making goals explicit, revisitable, and actionable throughout drafting and revision. Consistent with research showing that academic writing is guided by hierarchically structured and evolving goal systems [10,12], partic- ipants used goals not as fixed plans but as flexible reference points that were continuously refined as new ideas, sources, and feedback emerged, reflecting expert-like writing behavior [12]. Consistent with research showing that spe- cific, proximal, and process-oriented goals enhance metacognition and writing performance [22,19,5], participants reported WriteFlowâs emphasis on articulat- ing and refining goals helped them clarify vague intentions, focus on quality- oriented concerns, and maintain coherence across longer texts. Building on prior writing research showing that critically revising rather than passively accepting AI suggestions can foster critical thinking in academic writing [27], WriteFlow supported reflection on the underlying intentions of writers by linking goals to evolving text and prompting goal achievementâs evaluation. A key contribution relative to prior AI-supported writing systems lies in when metacognitive support is provided. Whereas tools such as VISAR focus on goal 12Y. Fan et al. construction during planning [29] and systems like Friction and ALure support reflection during revision [28,16], WriteFlow supports awareness of goalâtext alignment during drafting itself, addressing a gap where writers may otherwise fail to notice goal drift in AI-assisted writing. Features such as goal tracking and goalâtext linkage reduced the cognitive load of managing competing objectives, particularly during revision based on reviewer feedback. Participantsâ requests for hierarchical goal visualization and adaptive prioritization further highlight the value of making goal structures ex- plicit. This aligns with evidence that planning and quality-oriented goals sup- port higher-level revision [1], while underscoring the need to balance structure with cognitive load [18]. At the same time, participants noted limits of AI- supported evaluation: lengthy or overly authoritative feedback sometimes caused skepticism. Preferences for worked examples and multiple perspectives suggest that goal-oriented AI support is most effective when it scaffolds reflection and decision-making rather than prescribing solutions. 7.2 Enhanced Writer Agency and Critical Engagement with AI Addressing RQ2, the findings show that WriteFlow influenced the writerâs sense of agency by repositioning AI as a responsible, negotiable partner instead of an authoritative content generator. This is particularly salient given concerns raised in the formative study about overreliance, authorship, and ownership when using general-purpose AI tools (Section 3). Participants described WriteFlow as sup- porting their ability to preserve writer voice and meaning by requiring explicit decisions, such as accepting, rejecting, or revising goals and suggestions, at key points in the writing process. Participants contrasted this interaction style with experiences of using ChatGPT in a more passive or efficiency-oriented manner. In WriteFlow, agency was reinforced through goal-based framing of AI feed- back and through explanations that made the systemâs reasoning visible. This transparency enabled participants to critically evaluate AI outputs against their own intentions, supporting informed decision-making rather than deference. This aligns with participantsâ expectations for explainable AI and traceable reasoning, as articulated in design requirement 5 (R5). Importantly, verification and critique of AI outputs were not perceived as friction, but as productive moments of reflection. Participants reported that grounding AI feedback in concrete evidence from their own text increased trust while simultaneously encouraging scrutiny. Suggestions to visualize the AIâs in- ference process further indicate that transparency may serve not only trust cal- ibration, but also reflective engagement. However, participants also cautioned that single-perspective or overly polished AI outputs could constrain divergent thinking. Their preference for alternative options and multiple viewpoints un- derscores that preserving agency involves maintaining epistemic openness, not merely control. In this sense, WriteFlow was experienced as a dialogic partner that supported reflection-in-action [21], helping writers navigate uncertainty, evolving goals, and competing constraints in academic writing. From Intention to Text: AI-Supported Goal Setting in Academic Writing13 This study has several limitations and implications for the design of AI- supported academic writing tools. First, while the Wizard-of-Oz setup enabled fine-grained exploration of interaction dynamics, the use of researcher-prepared materials may have limited participantsâ sense of ownership. Future work should examine how goal-oriented AI support functions when writers engage with self- selected topics and authentic writing contexts over longer periods. Second, par- ticipantsâ feedback suggests that prompt design is not a purely interface-level concern, but a central mechanism shaping metacognitive engagement. Fixed prompt structures and uniform response lengths constrained reflective depth, indicating a need for adaptive prompting that responds to writersâ expertise, confidence, and stage in the writing process. Finally, the involvement of partici- pants with HCI expertise limits generalizability. Future studies should examine how writers from diverse disciplinary, linguistic, and educational backgrounds appropriate goal-oriented AI support, particularly given differences in writing conventions and self-regulatory practices. In sum, this work demonstrates how LLM-based writing tools can be designed to support metacognition and self-regulated learning by foregrounding iterative goal setting and goalâtext alignment throughout the writing process. Rather than optimizing for efficiency alone, WriteFlow illustrates a design direction in which AI systems scaffold reflective dialogue and evolving goal structures, reinforcing writer agency and intentional engagement in the academic writing process. Acknowledgments. We thank all the study participants for their engagement in this study. The work has in part been supported by the STINT grant: âCapitalizing on the potentials of technology to promote self-regulationâ (MG2018-7984). References 1. Beauvais, C., othersl: Why are some texts good and others not? relationship be- tween text quality and management of the writing processes. Journal of Educa- tional Psychology 103(2), 415 (2011). https://doi.org/10.1037/a0022545 2. Becker, J., Wahle, J.P., Gipp, B., Ruas, T.: Text generation: A systematic literature review of tasks, evaluation, and challenges. arXiv preprint arXiv:2405.15604 (2024). https://doi.org/10.48550/arXiv.2405.15604 3. Bentvelzen, M., WoĆșniak, P.W., Herbes, P.S., Stefanidi, E., Niess, J.: Revisiting reflection in hci: Four design resources for technologies that support reflection. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Tech- nologies 6(1), 1â27 (2022). https://doi.org/10.1145/3517233 4. Bereiter, C., Scardamalia, M.: An attainable version of high literacy: Approaches to teaching higher-order skills in reading and writing. Curriculum inquiry 17(1), 9â30 (1987). https://doi.org/10.1080/03626784.1987.11075275 5. Chung, H.Q., Chen, V., Olson, C.B.: The impact of self-assessment, planning and goal setting, and reflection before and after revision on student self-efficacy and writing performance. Reading and Writing 34(7), 1885â1913 (2021). https://doi. org/10.1007/s11145-021-10186-x 14Y. Fan et al. 6. Clarke, V., Braun, V.: Thematic analysis. In: Encyclopedia of critical psychology, p. 1947â1952. Springer, New York, NY, USA (2014). https://doi.org/10.1007/ 978-1-4614-5583-7_311 7. Dang, H., Benharrak, K., Lehmann, F., Buschek, D.: Beyond text generation: Sup- porting writers with continuous automatic text summaries. In: Proceedings of the 35th Annual ACM Symposium on User Interface Software and Technology. p. 1â13 (2022). https://doi.org/10.1145/3526113.3545672 8. E, J.L., et al.: When to give feedback: Exploring tradeoffs in the timing of design feedback. In: Proceedings of the 16th Conference on Creativity & Cognition. ACM, New York, NY, USA (2024). https://doi.org/10.1145/3635636.3656183 9. Fan, Y., Tang, L., Le, H., Shen, K., Tan, S., Zhao, Y., Shen, Y., Li, X., GaĆĄeviÄ, D.: Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance. British Journal of Educational Technology 56(2), 489â530 (2025). https://doi.org/10.1111/bjet.13544 10. Flower, L., Hayes, J.R.: A cognitive process theory of writing. College Composition & Communication 32(4), 365â387 (1981). https://doi.org/10.58680/c198115885 11. Freeman, J.: Student generative AI survey 2025. Tech. rep., Higher Edu- cation Policy Institute, London, UK (2025), https://w.hepi.ac.uk/reports/ student-generative-ai-survey-2025/ 12. Hayes, J.R., Flower, L.S.: Writing research and the writer. American psychologist 41(10), 1106 (1986). https://doi.org/10.1037/0003-066X.41.10.1106 13. Jansen, T., Meyer, J., Fleckenstein, J., Horbach, A., Keller, S., Möller, J.: Individ- ualizing goal-setting interventions using automated writing evaluation to support secondary school studentsâ text revisions. Learning and Instruction 89, 101847 (2024). https://doi.org/10.1016/j.learninstruc.2023.101847 14. Karolus, J., Feger, S.S., Schmidt, A., WoĆșniak, P.W.: Your text is hard to read: facilitating readability awareness to support writing proficiency in text production. In: Proceedings of the 2023 ACM Designing Interactive Systems Conference. p. 147â160 (2023). https://doi.org/10.1145/3563657.3596052 15. Li, T., et al.: Turning real-time analytics into adaptive scaffolds for self-regulated learning using generative artificial intelligence. In: Proceedings of the 15th In- ternational Learning Analytics and Knowledge Conference. p. 667â679 (2025). https://doi.org/10.1145/3706468.3706559 16. Neshaei, S.P., et al.: Leveraging learner errors in digital argumentation learning: How ALure helps students learn from their mistakes and write better arguments. Proceedings of the ACM on Human-Computer Interaction 9(2), 1â32 (2025). https: //doi.org/10.1145/3711023, https://dl.acm.org/doi/10.1145/3711023 17. Pineteh, E.A.: The academic writing challenges of undergraduate students: A south african case study. International Journal of Higher Education 3, 12â22 (09 2013). https://doi.org/10.5430/ijhe.v3n1p12 18. Radwan, A.Y., Alasmari, K.M., Abdulbagi, O.A., Alghamdi, E.A.: Sard: A human-ai collaborative story generation. In: International Conference on Human- Computer Interaction. p. 94â105. Springer (2024). https://doi.org/10.1007/ 978-3-031-76806-4_7 19. Reid, R., et al.: Strategy instruction for students with learning disabilities. Guilford Publications, New York, NY, USA, 2 edn. (2013) 20. Sabti, A.A., Md Rashid, S., Nimehchisalem, V., Darmi, R.: The impact of writ- ing anxiety, writing achievement motivation, and writing self-efficacy on writing performance: A correlational study of iraqi tertiary efl learners. SAGE Open 9, 215824401989428 (07 2019). https://doi.org/10.1177/2158244019894289 From Intention to Text: AI-Supported Goal Setting in Academic Writing15 21. Schön, D.A.: The reflective practitioner: How professionals think in action. Rout- ledge, London (2017). https://doi.org/10.4324/9781315237473 22. Schunk, D.H.: Goal setting and self-efficacy during self-regulated learn- ing. Educational psychologist 25(1), 71â86 (1990). https://doi.org/10.1207/ s15326985ep2501_6 23. TynjĂ€lĂ€, P.: Writing, learning and the development of expertise in higher education. In: Writing as a learning tool: Integrating theory and practice, p. 37â56. Springer, Dordrecht (2001). https://doi.org/10.1007/978-94-010-0740-5_4 24. Viberg, O., Khalil, M., Baars, M.: Self-regulated learning and learning analytics in online learning environments: A review of empirical research. In: Proceedings of the tenth international conference on learning analytics & knowledge. p. 524â533 (2020). https://doi.org/10.1145/3375462.3375483 25. Wong, J., Viberg, O.: Supporting self-regulated learning with generative ai: A case of two empirical studies. In: LAK Workshops. p. 223â229 (2024) 26. Xu, F., et al.: Are large language models really good logical reasoners? a com- prehensive evaluation and beyond. IEEE Transactions on Knowledge and Data Engineering (2025). https://doi.org/10.1109/TKDE.2025.3536008 27. Yang, K., RakoviÄ, M., Liang, Z., Yan, L., Zeng, Z., Fan, Y., GaĆĄeviÄ, D., Chen, G.: Modifying ai, enhancing essays: How active engagement with generative ai boosts writing quality. In: Proceedings of the 15th international learning analytics and knowledge conference. p. 568â578 (2025). https://doi.org/10.1145/3706468. 3706544 28. Zhang, C., et al.: Friction: Deciphering writing feedback into writing revisions through llm-assisted reflection. In: Proceedings of the 2025 CHI Conference on Hu- man Factors in Computing Systems. CHI â25, ACM, New York, NY, USA (2025). https://doi.org/10.1145/3706598.3714316 29. Zhang, Z., Gao, J., Dhaliwal, R.S., Li, T.J.J.: Visar: A human-ai argumentative writing assistant with visual programming and rapid draft prototyping. In: Pro- ceedings of the 36th annual ACM symposium on user interface software and tech- nology. p. 1â30 (2023). https://doi.org/10.1145/3586183.3606800 30. Zimmerman, B.J.: Becoming a self-regulated learner: An overview. Theory into practice 41(2), 64â70 (2002). https://doi.org/10.1207/s15430421tip4102_2 31. Zimmerman, B.J., Schunk, D.H. (eds.): Handbook of Self-Regulation of Learn- ing and Performance. Routledge, New York, NY (2011). https://doi.org/10.4324/ 9780203839010