Source-linked AI summary
Co-Writing with AI, on Human Terms: Aligning Research with User Demands Across the Writing Process
Mohi Reza, Jeb Thomas-Mitchell, Peter Dushniku, Nathan Laundry, Joseph Jay Williams, Anastasia Kuzminykh
TL;DR
How AI writing support shapes writers’ agency and ownership—and whether research aligns with their needs—remains an open question. This paper combines a systematic review with interviews, finding that preferred AI involvement varies by writing task: content-focused writers prioritize planning, while form-focused writers value control over translation and revision.
Problem
How AI writing research aligns with user needs and preserves writers’ agency and ownership across the writing process remains underexplored.
Source: §1 · p. 2Method
Results
Writers’ preferred AI involvement varies by task: content-focused writers emphasize control over planning and ideation, while form-focused writers value ownership in translation and revision.
Source: §7 · p. 27Takeaways & Limitations
Design guidance includes offering AI suggestions over direct edits, maintaining clear authorial boundaries, and providing global and local toggles to modulate AI involvement.
Source: §7 · p. 28The review’s focus on the ACM Digital Library may exclude valuable work from other venues and underrepresent interdisciplinary perspectives.
Source: §6.3 · p. 27
Abstract
from arXiv · showhide
As generative AI tools like ChatGPT become integral to everyday writing, critical questions arise about how to preserve writers' sense of agency and ownership when using these tools. Yet, a systematic understanding of how AI assistance affects different aspects of the writing process - and how this shapes writers' agency - remains underexplored. To address this gap, we conducted a systematic review of 109 HCI papers using the PRISMA approach. From this literature, we identify four overarching design strategies for AI writing support: structured guidance, guided exploration, active co-writing, and critical feedback - mapped across the four key cognitive processes in writing: planning, translating, reviewing, and monitoring. We complement this analysis with interviews of 15 writers across diverse domains. Our findings reveal that writers' desired levels of AI intervention vary across the writing process: content-focused writers (e.g., academics) prioritize ownership during planning, while form-focused writers (e.g., creatives) value control over translating and reviewing. Writers' preferences are also shaped by contextual goals, values, and notions of originality and authorship. By examining when ownership matters, what writers want to own, and how AI interactions shape agency, we surface both alignment and gaps between research and user needs. Our findings offer actionable design guidance for developing human-centered writing tools for co-writing with AI, on human terms.
1 Introduction
The paper examines how AI writing support can preserve writers’ agency and ownership while retaining AI’s broad usefulness, focusing on support across writing processes and alignment with user needs. A systematic review and interviews inform design guidance for human-centered collaboration.
1 Introduction
- AI writing tools can augment activities from brainstorming through revision, but their growing collaborative role raises concerns about writers’ agency and ownership.The paper frames abandoning AI as impractical because these tools support writers across the composition process.Source: §1 · p. 2
- The study addresses a gap in systematic understanding of writing processes and support strategies, and how they align with user perspectives across writing forms.Prior evaluations largely emphasized usability and efficiency, treating writing as one optimization task with limited attention to agency across distinct processes.Source: §1 · p. 2
- The research asks which design strategies support writing processes and contexts, and which processes writers want to control to preserve agency and ownership.It also examines how user situations, writing contexts, and AI interaction types shape perceptions of ownership.Source: §1 · p. 2
- The study reviews 1,676 Human-AI collaborative writing papers from 2018–2024 and interviews 15 writers, then synthesizes alignments and gaps between systems and expressed needs.The review classifies systems by composition processes using Flower and Hayes’s cognitive process theory.Source: §1 · p. 3
- The paper contributes four design strategies and actionable guidance for supporting writer agency and ownership across writing processes, grounded in when ownership matters and what writers want to own.It combines a systematic review of post-generative-AI research with user-centered analysis of writers’ ownership and originality concerns.Source: §1 · p. 3
2 Background & Related Work
The background establishes a cognitive-process lens for analyzing human–AI writing, traces the development of AI-assisted writing, and reviews evidence on agency and ownership to motivate a broad review of design strategies against user needs.
2.1 Theories on Writing Processes
- Flower and Hayes model writing as four dynamically coordinated processes: planning, translating, reviewing, and monitoring.Writers can move among processes and between high-level and local goals as they work.Source: §2.1 · pp. 3–4
- The study selects Flower and Hayes because its process model supports analysis of human–AI collaboration, unlike frameworks centered on writer–reader social relations such as Nystrand’s.Related work also examined planning, working memory, and social and physical environments, affect, and motivation.Source: §2.1 · p. 4
2.2 AI-Assisted Writing
- AI writing support developed from pre-transformer case-based story suggestions to transformer-era tools for creative, technical, accessibility, and specialized writing tasks.Early prototypes offered satisfying collaboration, but their text quality did not surpass unaided human writing.Source: §2.2 · p. 4
- Sentence-level suggestions promote original content while paragraph-level suggestions improve efficiency, and writers’ receptivity varies with confidence, expertise, goals, and values.Studies also report that biased models can influence users’ opinions and that professional writers struggle to maintain consistent style and voice.Source: §2.2 · pp. 4–5
2.3 Agency and Ownership in AI-Assisted Writing
- Greater AI intervention can undermine autonomy and control: longer completions reduced satisfaction, while increased support corresponded with lower perceived control.Autocomplete could threaten autonomy, whereas next-paragraph suggestions improved writing quality [29].Source: §2.3 · p. 5
- Ownership varies with writing purpose and contribution: expressive storywriters value direct control over translation, while idea generation can threaten writers who identify with overcoming writer’s block.Self-reported ownership also correlated with the proportion of user-written versus AI-generated text [74].Source: §2.3 · p. 5
- Perceived ownership strengthens when AI is viewed as enhancement rather than replacement, and depends on users’ contributions and leadership in the writing process.Related discussion identified involvement, infringement, and legal agreements as relevant ownership themes; the authors emphasize user control and balancing AI support with autonomy.Source: §2.3 · p. 5
- Previous HCI work had not broadly reviewed whether AI-writing design strategies preserve writers’ agency and ownership throughout the writing process.The review addresses this gap by evaluating researchers’ and designers’ strategies against users’ needs.Source: §2.3 · p. 5
3 Study 1: Reviewing Writing Process Dimensions in the Literature
Study 1 used a PRISMA systematic literature review and thematic coding to identify AI-writing design strategies and examine their distribution across writing processes and contexts.
3 Study 1: Reviewing Writing Process Dimensions in the Literature
- The review coded ACM Digital Library papers using Flower and Hayes’s writing-process theory and writing contexts, interfaces, and interactions enumerated by Lee et al. [73].Source: §3 · p. 6
- Thematic analysis identified strategies by interaction model, AI-support level, and treatment of outputs, then mapped systems by strategy across writing processes and contexts.Source: §3 · p. 6
3.1 Methods
The review followed PRISMA guidelines and examined AI writing-support research published from 2018–2024, the period associated with transformer-based language models’ emergence and widespread adoption.
3.1 Methods
- The review focused on papers published from 2018–2024, and used PRISMA guidelines to identify and analyze research on AI writing-support systems.The authors linked this period to the emergence and widespread adoption of transformer-based language models.Source: §3.1 · p. 6
3.1.1 Query Construction.
- The search query was iteratively refined from seed papers and expert knowledge, combining writing-related terms with AI-related terms to capture potentially relevant papers.The review included peer-reviewed English-language journal papers, conference proceedings, and extended abstracts.Source: §3.1.1 · p. 6, §3.1.2 · p. 6
3.1.2 Exclusion Criteria.
- The review excluded papers without AI-assisted writing support or user interaction, papers producing only code or images, and papers outside the specified contribution types.Eligible contribution types included user studies, artifacts or systems, theories or conceptual frameworks, and systematic reviews.Source: §3.1.2 · p. 6
3.1.3 Database Selection.
- In a preliminary comparison of 100 randomly sampled papers per database, ACM Digital Library had 11% relevant results versus 2% for IEEE Xplore and 3% each for Taylor & Francis and Wiley.The researchers selected ACM Digital Library as the most comprehensive and targeted corpus for their research questions.Source: §3.1.3 · pp. 6–7
3.1.4 Screening Process.
- The screening process yielded a final dataset of 109 papers after full-text review, with two researchers independently coding a 25-paper sample achieving Cohen’s kappa of 0.84 [82].The authors then used codebook thematic analysis, developing initial codes from Flower & Hayes’ cognitive process model and Lee et al.’s design space framework [73], with inductive themes informed by King & Brooks and Braun & Clarke.Source: §3.1.4 · p. 7, §3.1.5 · p. 7
3.2 Study 1 Findings
The review maps AI writing support across five contexts and identifies four strategies spanning writing processes, each with implications for agency, ownership, and task delegation.
3.2.1 Writing Context Characteristics.
- Academic systems emphasize structured skill development and user-owned planning and revision, whereas creative systems support ideation, blended generation, and feedback without direct revision.Research attention is uneven: Generating and Translating are strongly represented in Creative and Academic contexts, while Monitoring receives limited attention across contexts.Source: §3.2.1 · pp. 8–9, Table 1 · p. 8
- Formal support targets efficiency with limited evaluative reviewing; personal tools blend ideation and drafting as suggestions, while general tools enable rapid iteration and directly integrated revision suggestions.General-context systems also include accessibility support and are not adapted to a particular domain.Source: §3.2.1 · pp. 9–10
3.2.2 Strategies for AI-Assisted Writing Support in HCI Research.
- S1 scaffolds skill development through structured practice, iterative drafts, feedback, and separated workspaces that keep writers in control of textual changes.Its components include pattern mapping, sequential development, scaffolded feedback, and workspace control.Source: §3.2.2 · p. 10
- S2 guides iterative exploration and selection across an idea space, letting writers delegate generation while retaining control over which proposals enter the artifact.Its components include idea navigation, output variation, iterative revision, and proposal integration.Source: §3.2.2 · pp. 10–11
- S3 acts as an active writing partner for direct generation, content conversion, and efficient workflows, but integrating suggestions into final text can blur delegation boundaries and threaten agency.Users selectively offload writing while retaining editorial control over the final output.Source: §3.2.2 · p. 11
- S4 separates creation from analysis through structured feedback and manual integration, preserving writers’ control over whether and how to implement revisions.It specializes in reviewing and evaluation through qualitative feedback, quantitative analysis, hybrid evaluation, revision guidance, and analysis separation.Source: §3.2.2 · p. 11
- Among 62 systems, S3 led (23, 37.1%), followed by S1 (19, 30.6%), S2 (11, 17.7%), and S4 (9, 14.5%).Table 2 shows strategy distributions across writing processes and contexts; systems could be coded for more than one process or context.Source: Table 2 · p. 12, §3.2.2 · p. 12
- S1 dominated Evaluating (54.5%) and Academic writing (61.1%), while S3 led Generating (52.9%), Translating (50%), and Creative writing (50%); S2 led Formal writing (42.9%).The authors note that the single-coder approach is a limitation.Source: Table 2 · p. 12, §3.2.2 · p. 13
4 Study 2: Investigating AI’s Influence on Ownership in Writing
Study 2 used interviews with 15 writers to investigate which writing processes they consider essential to control for maintaining agency during AI-assisted writing.
4 Study 2: Investigating AI’s Influence on Ownership in Writing
- Interviews with 15 writers examined how user situations, writing contexts, and AI interaction types shape ownership perceptions across cognitive processes.The study asked which processes writers consider essential to control to maintain agency during AI-assisted writing.Source: §4 · p. 13
4.1 Methods
The study interviewed 15 writers with varied writing and AI experience, then analyzed their accounts and survey responses to examine ownership across distinct writing processes.
4.1.1 Participants.
- The study recruited writers with generative-AI expertise ranging from slightly to extremely knowledgeable, and all participants reported prior use of ChatGPT, with Grammarly and Microsoft Copilot next most popular.Sessions lasted 60–90 minutes online via recorded video calls; participants provided consent, could withdraw at any time, received CAN$20, and the institution’s research ethics board approved the protocol.Source: §4.1.1 · p. 13, §4.1.2 · p. 13
4.1.2 Procedure.
- Participants described their writing and AI experience in semi-structured interviews, then rated 16 Likert statements across the four writing processes.Sessions were recorded video calls; the post-interview survey followed a pre-survey and reflected on the discussion. The Likert ratings gauged preferences for AI involvement in each process.Source: §4.1.2 · pp. 13–14
- The researchers used reflexive thematic analysis, combining predefined writing-process codes with inductively identified patterns and keeping ownership aspects distinct.They reversed negatively worded survey items and used post-interview responses to quantify attitudes, without aggregating ownership and agency into a single measure.Source: §4.1.3 · p. 14
4.2 Study 2 Findings
Writers’ sense of ownership and willingness to delegate to AI shift with context, what they consider their primary contribution, and how the AI interaction preserves their agency. The findings distinguish content-focused ownership of ideas and planning from form-focused ownership of style and sentence-level choices.
4.2.1 When Ownership Matters.
- Time pressure and low-stakes tasks make writers more willing to delegate, while high stakes, limited confidence in specific skills, and distrust of AI shape selective delegation.Writers described using AI for polishing weaker areas without surrendering ownership of parts they considered central to their work.Source: §4.2.1 · pp. 15–17
- Writers value ownership most over work they see as their primary contribution, and may delegate tangential tasks when doing so preserves or enhances autonomy and creative joy.For content-oriented writers, that contribution often involves ideas; for form-oriented writers, it often involves style and expression.Source: §4.2.2 · p. 18
4.2.2 What Writers Want to Own.
- Content-focused writers prioritize owning ideas and planning, while form-focused writers seek control over style, translation, and revision; AI imitation of personal style can conflict with their values.A hobbyist fiction writer, by contrast, delegated initial worldbuilding to AI to move directly to sentence writing.Source: §4.2.2 · pp. 18–19
4.2.3 How AI Interactions Shape Ownership.
- Writers preferred optional AI suggestions over direct insertion or overwriting, and maintained ownership by retaining final say and consciously selecting or refining suggestions.They framed AI as inspiration rather than substance for primary contributions, while allowing more selective use for less critical tasks.Source: §4.2.3 · pp. 19–20
- Global toggles let writers manage suggestions around focus and flow, while local controls let them reject AI corrections that would disrupt intentional stylistic rule-breaking.Writers connected control over AI input to preserving both concentration and an authentic voice.Source: §4.2.3 · p. 20
5 Alignment Between the Two Studies · 5.1 Alignment with Contextual Factors of Ownership
The comparison examines how HCI writing-support strategies align with writers’ ownership concerns, including how context shapes willingness to delegate. Existing research addresses time constraints, trust, task importance, and perceived competence, while the analysis emphasizes considering all four together.
5 Alignment Between the Two Studies
- Section 5 compares literature-review findings with interviews across contextual factors, writing-process preferences, and interaction choices that shape writers’ agency.The supplied subsection develops the contextual-factors dimension.Source: §5 · p. 20
5.1 Alignment with Contextual Factors of Ownership
- Writers’ ownership concerns depend on time constraints, trust, task importance, and perceived competence, dimensions that existing HCI research has investigated extensively.These factors reflect personal values and external limitations that shape willingness to delegate writing tasks to AI.Source: §5.1 · p. 21
- In collaborative writing, Shakeri et al. found that offloading narrative tasks to AI helped alleviate time pressures, addressing time constraints as a concern in collaborative contexts.Source: §5.1 · p. 21
- Professionals’ trust in AI collaborators depends on explainable, actionable feedback and shared social context.These requirements reflect preferences similar to those of the interview participants.Source: §5.1 · p. 22
- AI-tool use varies between professional and non-professional users, while creative writers may feel vulnerable when they lack confidence [113].The passage identifies perceived competence as a contextual factor affecting ownership concerns.Source: §5.1 · p. 22
- A multi-level framework lets writers choose AI creative involvement for different task stakes; this analysis argues that all four contextual factors should be considered together.Prior studies addressed one or two factors, while the analysis proposes a foundation for research considering their full range rather than in isolation.Source: §5.1 · p. 22
5.2 Alignment with Essential Cognitive Processes · 5.3 Alignment with Desired Interfaces and Interactions
Writers’ ownership priorities vary by writing context, and no AI design strategy perfectly matches their desired delegation across writing processes. Interface designs can support agency through suggestions and editorial control, but fine-grained control over AI involvement remains underexplored.
5.2 Alignment with Essential Cognitive Processes
- No AI design strategy maps perfectly onto writers’ differing ownership demands across form-centric and content-centric contexts, motivating flexible control over AI involvement across writing processes.The study distinguishes contexts by what writers want to own and reports this mismatch in Table 4.Source: §5.2 · p. 22
- Form-centric writers prioritize ownership of translation and revision while welcoming AI planning, yet all strategies offer more translation and review support than they want.Active Co-writing is the plurality in Creative, Personal, and General contexts; exploratory Guided Exploration methods prompt writers to develop ideas themselves, and work such as merits further investigation.Source: §5.2.1 · p. 22
5.2.1 Form-centric Writers.
- Content-centric writers prioritize generating and organizing ideas and setting goals, while S1 and S2 align with their desired translation, evaluation, and revision support.S1 and S2 comprise 61% of Academic-context systems and 57% of Formal-context systems, but provide more planning support than these writers want, suggesting a need for granular control.Source: §5.2.2 · p. 22
5.3 Alignment with Desired Interfaces and Interactions
- Presenting AI content as suggestions is common and aligns with user demands; suggestion length is inversely associated with perceived text ownership.Research has examined visual differentiation of suggestions [8] and suggestion length or quantity [29] [36].Source: §5.3.1 · pp. 22–23
5.3.1 Suggestions.
- Workspace suggestions can blur user–AI boundaries, whereas separated suggestions reinforce the writer’s decision-making role and control over final text.Some systems place suggestions directly in the workspace [8] [18], while others use a separate interface [29] [43] [89].Source: §5.3.1 · p. 23
- All four design strategies preserve the writer’s editorial control in the dataset, but experimental conditions removing user influence over AI text were associated with reduced perceived ownership.Workspace Control (S1) requires explicit action, Proposal Integration (S2) offers suggestions, Result Ownership (S3) streamlines integration with possible authorship tension, and Analysis Separation (S4) requires manual integration; the no-influence conditions are reported in.Source: §5.3.2 · p. 23
5.3.2 Final Say.
- Global and Local AI Toggles are absent from systems in the dataset, and user-initiated interactions alone do not provide the fine-grained control writers want.The authors encourage systems that let writers modulate collaboration, restrict assistance to designated components, or remove it altogether; cited theoretical frameworks include and.Source: §5.3.3 · p. 23
5.4 Monitoring
Monitoring is underexplored in AI writing research, even though writers must oversee both their own writing and AI contributions. Flexible controls over AI involvement could support writers’ nuanced relationship between assistance and ownership.
5.4 Monitoring
- Monitoring is significantly underexplored, although AI makes it more complex by requiring writers to oversee both their own writing and AI contributions.Monitoring is distinct from the human–AI collaborative relationship but connects to managing both individual and collaborative writing processes.Source: §5.4 · p. 23
- Research may overlook broader collaborative dynamics by optimizing specific interactions; for example, suggestion studies do not let participants disable AI assistance entirely.The authors identify this as a likely reason monitoring has received less attention.Source: §5.4 · p. 23
- Flexible interfaces that let writers modulate AI involvement could better support dynamic control and the nuanced relationship between assistance and ownership found in interviews.The authors frame monitoring and management of the human–AI relationship as an increasingly important research opportunity for CSCW.Source: §5.4 · pp. 23–24
6 Discussion
The discussion synthesizes a process-specific account of ownership with a map of AI-writing strategies, offering guidance for aligning assistance with writers’ valued contributions. It presents the combined review and interview approach as a way to inform agency-centered design.
6.1.1 RQ 1.1 Design Strategies .
- The review identified Structured Guidance, Guided Exploration, Active Co-Writing, and Critical Feedback, supporting skill-building, controlled exploration, efficient collaboration, and reflection, respectively.Examples include step-by-step support for literature reviews, iterative exploration of text, selecting AI suggestions, and audience-specific feedback.Source: §6.1.1 · p. 24
6.1.2 RQ 1.2 Process & Context Distribution.
- S3 comprised 37.1% of all systems and 50% of Creative-writing systems, while S4 comprised 14.5% of the dataset; Monitoring received limited support across contexts.The authors suggest the prevalence of S3 in Creative writing may misalign with ownership needs because active co-writing can compromise ownership central to creative expression.Source: §6.1.2 · p. 25
- Content-focused writers prioritize planning and original ideas, whereas form-focused writers protect control over translation, revision, voice, tone, and word-level choices.For content-focused writers, delegating planning to AI can undermine ownership; for form-focused writers, control of stylistic choices is central to their identity as writers.Source: §6.1.3 · p. 25
6.1.3 RQ 2.1 Agency-Preserving Processes:
- Ownership is flexible, shaped by time, importance, confidence, trust, content-versus-form priorities, and interaction options including suggestions, final say, and global or local AI toggles.These factors shape when ownership matters, what writers want to own, and how they experience agency during AI-assisted writing.Source: §6.1.4 · p. 25
6.1.4 RQ 2.2 Modulating Factors in Ownership:
- Writers seek AI help outside their primary contribution, with Table 3 distinguishing preferred support, areas where AI should not intervene, and zones where it may assist cautiously.This “chessboard-like” pattern links content-focused writers’ ownership to planning and form-oriented writers’ ownership to translation and review.Source: §6.1.4 · p. 25
6.2 Contributions to CSCW
- The study adds process- and context-level granularity to AI-writing design research and finds that style mimicry can feel invasive to form-oriented and expert writers.It relates to CSCW’s interest in human–AI collaboration and extends prior mapping work [73]; its multidimensional view of automation and control resonates with Shneiderman’s HAI framework and contrasts with a unidimensional account.Source: §6.2 · p. 26
6.3 Limitations and Future Work
- Interpret the findings cautiously because the theoretical lens, ACM-only review, and narrow interview sampling constrain completeness and generalizability.The sample was limited in age (18–34), education, and prior AI familiarity; recruitment was by convenience, and participants’ tool experience centered on ChatGPT. The authors also note that the framework may not fully describe creative and professional Human–AI interaction, and that other libraries could broaden the literature review.Source: §6.3 · pp. 26–27
7 Conclusion
The review and writer interviews show that preserving agency in AI-assisted writing requires adapting AI involvement to writing processes, tasks, and context. The paper offers design guidance for co-writing that keeps users in control.
7 Conclusion
- Content-focused writers prioritize control over planning and ideation, while form-focused writers value ownership in translation and revision.Preferences also depend on task and contextual factors including time pressure, trust, task importance, and perceived competence.Source: §7 · pp. 27–28
- The review identifies four AI writing design strategies: structured guidance, guided exploration, active co-writing, and critical feedback.Source: §7 · p. 27
- Adaptive writing systems should favor AI suggestions over direct edits, maintain clear authorial boundaries, and offer global and local toggles for AI involvement.These recommendations aim to align system design with writers’ real-world needs and preserve user agency.Source: §7 · p. 28