AI writing assistants have been studied a lot — but which combos of user, task, and style have researchers never actually tried?

This explores where the open gaps are in AI writing assistant design: which pairings of user, task, interaction style and underlying model researchers haven't yet built or tested. The honest answer starts with a limit. The best map of the field is a review of 115 papers that sorts writing assistants along five aspects (task, user, technology, interaction and ecosystem), split into 35 dimensions (What dimensions structure the design of AI writing assistants?). That map was built to show which combinations have been tried and which haven't. But the corpus here describes the map, not its list of empty cells. So what follows is an inference from where individual studies cluster and where they stop, not a definitive inventory.

One pattern stands out. Most studies vary one aspect at a time, and the interesting gaps sit where two of those threads have never met. For example, writers can usefully set up a proactive AI partner ahead of time, choosing its role and how often it should speak up (Can writers benefit from configuring AI writing partners in advance?). Separately, writers in shared editors want to see when and how their co-authors prompted the AI (Do writers want to see each other's AI prompts in shared editors?). Nobody here has studied a proactive partner that a whole team configures together. Whose settings would win? Similarly, writers move between ideation, organizing their thoughts and drafting, and they loop back to ideation when they get stuck (How do writers use AI through different creative stages?). Yet the setup study treats proactivity as one fixed choice rather than something that should change with the stage.

The second gap is about what the assistant is optimized for. Writers prefer AI rewrites 63% of the time, but those same rewrites shift how readers see the writer on all 29 traits measured. Writers come across as more confident, more extreme and more privileged (Does AI writing assistance change how readers perceive the writer?, Can user preference guide AI writing tool alignment?). The polish and the distortion come bundled together. On top of that, GPT-4o autocomplete pulled Indian writers toward Western phrasing and gave Americans bigger productivity gains (Do AI writing assistants push non-Western writers toward Western styles?). Put those together and you get a largely empty cell: assistants built to protect a writer's own voice and cultural style rather than to maximize preference ratings. A related idea comes from outside writing research. Giving a model an explicit list of what it doesn't yet know about the user cut sycophancy and hallucination roughly in half (Do language models know what they don't know about users?). No writing tool here applies that idea to a writer's voice.

The third gap is the interface itself. Nielsen estimates that only 10 to 20 percent of adults in wealthy countries write well enough to prompt AI effectively (Can most adults write prompts good enough for AI?). Friedman lists alternatives to typing prompts, such as style lenses, visual maps of intent and direct manipulation on a canvas (Can task-oriented UI patterns reduce the friction of AI chat interfaces?). The irony is that the writing-assistant studies here are still mostly prompt-driven, which assumes users who are good at writing prompts. There's also a quieter structural gap. ChatGPT tends to recap what it has already said, while human writers preview what's coming next (Does ChatGPT organize text differently than human writers?). Assistants that help with document-level structure, rather than sentence-level polish, barely show up.

The takeaway you might not have expected: the biggest open territory isn't a new feature. It's the set of combinations that cross aspects of the map, such as social settings with proactive behavior, voice preservation with cultural fit, and non-chat interfaces with less fluent writers. The individual pieces exist, but nobody here has put them together.