A Design Space for Intelligent and Interactive Writing Assistants

1 INTRODUCTION

• The design space addresses fragmentation across NLP, HCI, CSS, and more specialized communities, which study writing assistants with different emphases and can struggle to keep up with work elsewhere.The paper defines writing assistants broadly as computational systems that improve writing quality and effectiveness, from checking grammar to generating ideas and restructuring text.Source: §1 · p. 1
• It aims to provide a shared vocabulary, support analysis of existing designs, and help stakeholders envision designs in unexplored regions of the design space.The paper also offers two illustrative scenarios demonstrating use by stakeholders including researchers and policymakers.Source: §1 · pp. 1–2
• The design space organizes writing assistants across five aspects—task, user, technology, interaction, and ecosystem—with 35 dimensions and 143 codes derived from a systematic review of 115 papers.Its scope includes writing contexts, user relationships such as agency and trust, technology learning problems, and ecosystem factors such as infrastructure, norms, and change over time.Source: §1 · pp. 1–2
• The authors present the framework as a practical, non-exhaustive and non-permanent resource for exploring, understanding, and comparing writing assistants beyond researchers’ and designers’ immediate fields.They publicly release annotated papers as a living artifact to support community refinement and extension of the design space.Source: §1 · p. 2

• Writing technologies progressed from cuneiform and typewriters to word processors with flexible editing and support for writers’ cognitive processes.Examples of cognitive support include text analysis and text planning.Source: §2 · p. 2
• Cognitive support may improve writing proficiency and efficiency, but persistent support, interaction choices, fading scaffolds, and complex-text understanding remain challenges as NLP language models advance [22] [186].One research thread studies the human writing process, including cognitive process theory, while another extends assistant intelligence to handle long, complicated texts.Source: §2 · p. 2
• Recent assistants target domain-specific tasks [85], while studies examine effects on behavior, human–LM interaction, prompting [58], and steering through visual sketching.Examples include help-request writing, affectionate instant messages, and scientific tweetorials; studies also investigate how people interact with and receive support from these systems.Source: §2 · p. 2
• Despite growing ethical and practical concerns, existing analyses cover only selected dimensions, and no comprehensive study has mapped the writing-assistant landscape.Concerns include students’ AI use in writing and homework; prior analyses examine interaction patterns, writing goals and technologies, and user perspectives, while creativity-support spaces have been mapped.Source: §2 · p. 2

3.1 Scope

• The scope defines intelligent assistants as capable of autonomous decisions or text generation, interactive systems as involving human input or output and iteration, and writing as producing natural-language text.These are working definitions, not universal ones; programming, multi-user collaborative, and text-entry tools are excluded.Source: §3.1 · pp. 2–3

3.2 Five Aspects of Writing Assistants

• The design space treats writing assistants as five interdependent aspects—task, user, technology, interaction, and ecosystem—with dimensions and codes describing components and possible options.The aspects cover writing goals and contexts, user characteristics, technical building blocks, interaction paradigms, and broader economic, social, and regulatory conditions.Source: Figure 1 · p. 3, §3.2 · p. 4

3.3.1 Paper Selection.

• The review screened 419 HCI/NLP venue papers retrieved through August 2023 and selected 115 explicitly connecting technology, a human user, an interface, and writing.Three authors independently annotated relevance, discussed disagreements, and refined the scope before selecting papers.Source: §3.3.1 · p. 4

3.3.2 Code Development.

• Five aspect-based teams iteratively reviewed papers with at least two readers and refined codes through discussion and recoding as needed.Code development combined inductive coding from reviewed papers with deductive codes informed by external knowledge; the technology team also consulted 25 additional papers not included in the review.Source: §3.3.2 · pp. 4–5

3.3.3 Final Design Space and Coding.

• The final pass coded all 115 papers with 35 dimensions and 143 codes, with Percent agreement (mean: 0.93, std: 0.06) and Krippendorff’s alpha (mean: 0.69, std, 0.17).The authors released the coded papers as a living artifact; codes occurring fewer than 50 times were excluded from the alpha calculation.Source: §3.3.3 · p. 5

4.1 Task

• The framework treats writing tasks as user-identified rhetorical purposes at particular writing stages and in specific contexts, with varying requirements and intended audiences.This user-centered definition makes task objectives central to how an assistant is designed.Source: §4.1 · p. 5

4.1.1 Dimensions and Codes.

• Task dimensions cover writing stage, context, purpose, specificity, and audience; stages group related cognitive processes, with multiple processes often involved in one stage.For example, planning an outline can involve planning, translating ideas into bullets, and reviewing their flow.Source: §4.1.1 · pp. 6–7

4.2 User

• The user aspect distinguishes who may benefit from an assistant and emphasizes that effectiveness depends on attunement to users’ differing needs and preferences.Source: §4.2 · p. 7

4.2.1 Dimensions and Codes.

• User dimensions include demographics, capabilities, system relationships, and output preferences, spanning attributes such as expertise, agency, trust, privacy, coherence, diversity, explainability, bias, and personalization.Capabilities include writing efficiency, creativity, emotion, empathy, cognition, and neurodiversity; relationships include ownership, integrity, and transparency.Source: §4.2.1 · p. 9

4.3 Technology

• The technology aspect examines the building blocks that underpin assistants’ intelligence and capabilities: data, models, learning problems and techniques, and evaluation.Source: §4.3 · p. 9

4.3.1 Dimensions and Codes.

• Technology dimensions characterize data by source and size, models by type and external-resource access, and learning by problem formulation, algorithm, and training or adaptation approach.The passage distinguishes rule-based, statistical, neural, and foundation models; learning problems include classification, regression, structured prediction, rewriting, generation, and retrieval.Source: §4.3.1 · pp. 11–12
• Evaluation dimensions identify who assesses model outputs, including automatic, machine-learned, human, and hybrid human-machine evaluation.Hybrid evaluation is especially relevant to co-creative, mixed-initiative writing assistance.Source: §4.3.1 · pp. 12–13
• Evaluation spans linguistic quality, controllability, contextual style and adequacy, and ethics, while scalability concerns development cost, inference cost, and latency.Prompting or efficient fine-tuning can avoid full fine-tuning costs; quantization and distillation can address deployment costs and latency.Source: §4.3.1 · p. 13

4.4 Interaction

• Interaction is organized around three components: the user, the interface as mediator, and the system, with users providing input and perceiving system outputs through the interface.The system accesses user input and data and generates outputs rendered in the interface.Source: §4.4 · p. 13, Figure 3 · p. 14

4.4.1 Dimensions and codes.

• Interface dimensions cover agent, tool, and hybrid metaphors; writing-area, separated, embedded, or custom layouts; editor, chatbot, or custom paradigms; and ways of visually distinguishing system output.Visual differentiation may use formatting or a separate panel, may be unnecessary for different media, or may be absent when generated text is inserted identically to user text.Source: §4.4.1 · pp. 13–14
• Interaction dimensions distinguish user- from system-initiated output, implicit from explicit steering or no control, and user choices to select, retype, edit, or omit system output.Implicit steering through the artifact can be less disruptive, while explicit controls offer more ways to express intentions.Source: §4.4.1 · p. 14
• System dimensions cover accessible user data, output types such as analysis, generation, and proposals, and curation through model-generated, pre-curated, or deterministic responses.Additional user data can include explicit task instructions or sketches; pre-curated responses trade model flexibility for greater control and may be dynamically customized.Source: §4.4.1 · pp. 14, 16

4.5 Ecosystem

• The ecosystem is the sociotechnical context surrounding writers and tools, including interdependent actors beyond designers’ immediate control that designers can anticipate.Its dimensions include infrastructure compatibility and interoperability, openness and licensing, stakeholders and social support, physical locale, and norms and rules.Source: §4.5 · pp. 16–17, §4.5.1 · pp. 17–18

4.5.1 Dimensions and codes.

• Ecosystem design also considers formal and informal norms and rules, while change over time encompasses evolving writers, readers, technologies, regulation, and information environments.The corpus included empirical studies of immediate writing effects but no studies beyond a single writing session or longitudinal studies of readers and reading practices; the authors identify longer-term risks and adaptation needs.Source: §4.5.1 · p. 18

5.1 Use Case Scenarios for the Design Space

• A generative scenario shows how mapping design choices can reveal overlooked decisions and prompt a richer assistant design.Researchers designing paraphrase support for non-native English writers identify gaps around model data and evaluation, trust, transparency, and interaction metaphor, building on a hypothetical scenario based on Kim et al. [136].Source: §5.1 · p. 19
• An analytic scenario shows how mapping a study can help policymakers identify factors that may shape findings and inform possible safeguards.In a hypothetical scenario based on Jakesch et al., policymakers consider system-initiated suggestions and implicit steering toward a predetermined stance, then consider user-initiated interaction, explicit steering, and visual differentiation of AI text.Source: §5.1 · p. 19

5.2 Trends and Gaps in the Literature

• 13 papers used the foundation-model code in 2023 versus 1 in 2020, amid a mid-2010s rise and underrepresentation of audience, scalability, and ecosystem dimensions.The rise was roughly equal across HCI and NLP venues, with slightly more HCI papers. The authors also note limited attention to trust, transparency, controllability, and ethics as foundation-model use grows.Source: §5.2 · pp. 19–20, Figure 4 · p. 19, Figure 5 · p. 20

5.3 Ethical Implications of Writing Assistants

• Writing assistants can enable plagiarism, deceptive content, and opinion swaying, while raising labor-market and equity concerns.The authors call for ethical governance attentive to sector-specific impacts and diverse users’ needs, preferences, and capabilities [1] [14], especially in education, counseling, and healthcare.Source: §5.3 · p. 20

5.4 Challenges in Developing a Design Space

• Developing the design space involved blurry aspect boundaries, dimensions that were sometimes implicit, and differences in how aspects classify or record codes.The authors duplicated user dimensions to code both design choices and reported properties, while retaining an official unduplicated set; they also had to discretize continuous codes, select frequently observed options, and trade specificity for generalizability.Source: §5.4 · p. 21

5.5 Limitations and Future Work

• The corpus may omit relevant papers and commercial assistants, and coding may misinterpret authors’ intentions; future work will refine the space and code commercial systems.The search criteria focused on “write” in titles or keywords, and the authors prioritized useful dimensions and codes over exhaustive paper coverage.Source: §5.5 · p. 21

5.6 Lessons for Future Writing Assistants

• The authors offer the design space as a non-prescriptive reference for navigating trade-offs and encourage researchers to explore underrepresented dimensions as technology and user needs change.They anticipate new capabilities, interaction designs, and insights about user preferences and behavior, and urge designers to remain adaptive.Source: §5.6 · p. 21

6 CONCLUSION

• The design space organizes writing assistants across five aspects using 35 dimensions and 143 codes, offering a practical tool for comparison and design.It was developed through community collaboration and systematic literature review across multiple disciplines.Source: §6 · p. 21

A AUTHOR CONTRIBUTIONS

• The project involved 36 researchers across 27 institutions, with author contributions documented in three ways to support appropriate acknowledgment and future inquiries.The paper and artifact list authors by overall contributions, team-specific contributions, and core annotation or artifact work.Source: A AUTHOR CONTRIBUTIONS · p. 30

A.1 Overall Author List

• Authors initially took on project lead, team lead, team member, or advisor roles, with some authors serving in two roles.Project leads oversaw the project; team leads coordinated reviews and writing; team members reviewed literature and wrote; advisors provided guidance and feedback.Source: §A.1 · p. 30

A.2 Team-Specific Author Lists

• The team-specific author lists divide contributions across task, user, technology, interaction, and ecosystem teams, naming leads, members, and advisors for each aspect.Task and user team rosters appear separately from the technology, interaction, and ecosystem rosters.Source: §A.2 · pp. 31–32

A.3 Core Group of Annotators

• After teams developed and iterated on their own dimensions and codes, a core group volunteered to annotate all papers across all dimensions, reducing communication overhead.Most authors initially coded a relevant subset of papers for their team; the smaller core group then coded beyond its members’ team assignments.Source: §A.3 · p. 32
• The paper names 15 core annotators and identifies Shannon Zejiang Shen and Mina Lee as artifact designers.The annotator list is ordered by contribution, with the first person having annotated the highest number of papers.Source: §A.3 · p. 32

B TERMINOLOGY

• Writing assistants are computational systems that assist users with writing and require a frontend for interfacing with users.Source: B TERMINOLOGY · p. 32
• A model means a specific model instance, while system broadly covers writing assistants, models, model-plus additions, and non-model-based systems.A model-plus system can combine a model with safety filters or external resources.Source: B TERMINOLOGY · p. 32
• Technology is broader than model or system, encompassing data, models, learning, evaluation, and beyond.Source: B TERMINOLOGY · p. 32

C.1 Technological Evolution in Writing Assistants

• Model architectures and the amounts and types of data they could use evolved together in writing assistants.Source: §C.1 · p. 32
• Early-2000s systems used human-labeled examples numbering in the hundreds or thousands to develop rule-based methods or train statistical models.Source: §C.1 · p. 32
• As statistical machine learning and deep neural networks advanced [172], writing tools could use larger unlabeled sources, including learned or task-trained word embeddings.Source: §C.1 · p. 32
• Word embeddings enabled systems to bootstrap from unlabeled data while using smaller human-labeled datasets for further modeling and evaluation.Source: §C.1 · p. 32
• Deep sequence-to-sequence models [42] supported writing assistants trained with human- and machine-labeled data comprising several thousand sentences or examples.Source: §C.1 · p. 32
• Recent systems use large language models pretrained on corpora of millions of sentences [35] [58] [65], then tune them with small labeled datasets or instruction tuning [204] [218] [226].Source: §C.1 · p. 32
• Some recent systems instead prompt pretrained models zero-shot to perform writing tasks.Source: §C.1 · p. 32

C.2 User Interaction & Interface Evolution in Writing Assistants

• AAC systems used next-word prediction to reduce typing, and later systems applied these ideas more broadly as technology advanced from n-grams to deep learning and prompting.These advances enabled more coherent, longer text generation and influenced interaction and UI design.Source: §C.2 · p. 32
• Generated output expanded from single words and phrases to whole drafts, while users could steer completion implicitly through draft text or explicitly with instructions.Explicit instructions are often paired with chat-history interfaces rather than traditional page views.Source: §C.2 · p. 32
• Interfaces present AI text in chat histories, sidebars, separate views, inline previews, cursor pop-ups, sketches, and two-dimensional canvases.Canvas formats include graphs and post-it-note arrangements.Source: §C.2 · pp. 32–33

C.3 Ecosystem: Going from Micro-HCI to Macro-HCI

• Historical accounts trace HCI’s scope from the computer and individual user to wider sociotechnical systems shaped by interacting groups and interdisciplinary perspectives.Source: §C.3 · p. 33
• Shneiderman [225] distinguishes micro-HCI’s interface design for varied devices and users from macro-HCI’s broader concerns, open tasks, unanticipated goals, and community conflicts.Macro-HCI also encompasses affective experience, aesthetics, motivation, social participation, trust, empathy, responsibility, and privacy.Source: §C.3 · p. 33
• Task, user, technology, and interaction describe micro-HCI, while ecosystem extends the design space to macro-HCI concerns beyond the individual writer and software.Source: §C.3 · p. 33

D.1 Venues

• The review included all paper tracks at its selected venues, such as CHI Late-Breaking Work and ACL Findings.Workshops were excluded.Source: §D.1 · p. 33
• The selected HCI venues were CHI, CSCW, UIST, IUI, C&C, DIS, and ToCHI.Source: §D.1 · p. 33
• The selected NLP venues were ACL, NAACL, EMNLP, EACL, and TACL.Source: §D.1 · p. 33

D.2 Keywords

• For ACM DL, the review searched for “write,” which automatically matched variations, and retrieved papers with the keyword in their title or keywords.Source: §D.2 · p. 33
• For ACL Anthology, the review searched for “writ” and “wrote” to account manually for variations including “writing” and “written.”Source: §D.2 · p. 33
• Because ACL Anthology papers lacked associated keywords, the review retrieved papers with at least one search term in the title, finding 60 HCI papers and 55 NLP papers.Source: §D.2 · p. 33

D.4 Additional References for Technology

• The team selected 25 technology papers from an initial set of 80, using expanded keywords and venues, deduplication, and relevance review by two authors.The selected papers were those both authors considered relevant to the technology aspect.Source: §D.4 · p. 33
• Some selected papers were out of scope for the literature review but remained relevant as AI models for writing tasks, and the team used the papers to develop codes.Examples included language models fine-tuned for text composition or revision; the papers were not included in the literature review.Source: §D.4 · p. 33