What dimensions structure the design of AI writing assistants?

Through "a large community collaboration," the authors propose "a design space as a structured way to examine and explore the multidimensional space of intelligent and interactive writing assistants." They systematically review 115 papers and employ an iterative coding process to arrive at "five key aspects of writing assistants—task, user, technology, interaction, and ecosystem," within which they identify "35 dimensions ... and 143 codes." The aspects named in the excerpt include task ("writing stages, contexts, and purposes"), user ("characteristics and preferences of users"), and technology ("building blocks of developing underlying models"); interaction and ecosystem are named as the two remaining aspects but not defined in the excerpted text.

The authors derive their structure from the sociotechnical systems perspective, which treats "task," "user," "technology," and "structure" as "integral and interdependent parts of a complex system." They adapt it by splitting "technology" into "technology" and "interaction" "to be more aligned with current research in NLP and HCI, respectively," and renaming "structure" to "ecosystem" to capture "the broader context in which writing assistants operate." They scope "intelligent" as capable of "autonomous decision-making and/or text generation," "interactive" as reflecting "human input and/or output" in "an iterative process," and exclude programming tools, collaborative multi-user tools, and text-entry tools from their 115-paper sample. They argue design spaces serve three uses: establishing "a shared vocabulary," supporting "understanding existing designs," and supporting "envisioning new designs" by surfacing dimension combinations that "no design considers" yet.

Two neighboring notes read as instances landing inside specific cells of this space. Can writers benefit from configuring AI writing partners in advance? studies a writing assistant where relevant partners take the initiative at appropriate moments — a concrete instance of the dimension the authors' own analytic scenario calls "UI - initiative [system-initiated]" versus user-initiated. Do writers want to see each other's AI prompts in shared editors? empirically tests what the authors' analytic scenario names "UI - visual differentiation" and the user aspect's "relationship to system [trust, transparency]" — the authors even recommend, from their scenario, that "designers visually differentiate between user-generated and AI-generated text," which is the exact design choice the prompt-sharing study varies and measures preferences over. The design space supplies the vocabulary; these two papers fill named cells of it with findings.

The excerpt is a classification scheme, not an evaluation: it does not claim that any particular combination of dimensions and codes produces a better writing assistant, only that the dimensions exist in the literature. The authors flag their own limits — "write" in titles or keywords may undercount relevant NLP papers, commercial writing assistants without research papers were excluded, and coding errors from "misinterpreted authors' intentions" are possible. What follows is narrower than a map of what's good: it is a map of what has been tried and named, useful for spotting gaps (as in the paper's own generative scenario, where researchers find they "overlooked" dimensions like data source and evaluation focus) and for auditing a specific study's design choices (as in the analytic scenario), but silent on which choices writers are better served by.