What design changes could reduce unhelpful AI reliance in collaborative writing tools?

This explores what we could redesign in AI writing tools so people lean on them less passively — and the corpus suggests the problem isn't that AI writes badly, but that it writes persuasively enough to be accepted unexamined. The starkest finding: writers edited AI-generated paragraphs only 23% of the time, and when they did, their edits stayed 96% similar to the original Do writers actually edit AI-generated text before publishing?. That near-total acceptance matters because the same assistance systematically distorts the writer's persona — across all 29 measured dimensions, pushing text toward more confidence, more agreeableness, more perceived privilege Does AI writing assistance change how readers perceive the writer?. So the reliance is unhelpful in a specific way: it smuggles in a voice that isn't the writer's, and the writer ships it unread.

The tempting design fix — just optimize for what writers prefer — turns out to be a dead end. Writers prefer AI rewrites 63% of the time, yet object to the persona distortions those very rewrites introduce, and the polish and the distortion are entangled at the model level: you can't optimize for one without producing the other Can user preference guide AI writing tool alignment?. This reframes the design target. If preference-following both drives over-reliance and bakes in distortion, then a tool tuned to please is structurally part of the problem. The lever is friction and transparency, not smoother suggestions.

Where should that friction go? A study of how writers actually use AI across creative stages found they reach for it most heavily at ideation, then organizing, then drafting — and that unexpected outputs trigger genuinely new directions How do writers use AI through different creative stages?. That's a design opening: tools could lean into AI as an idea-divergence engine (where surprise is the value) while deliberately raising the cost of accepting finished prose verbatim (where the 23% edit rate does its damage). Stage-aware assistance — generous early, resistant late — fits the way people already work better than a uniform autocomplete that's equally eager at every stage.

The deeper diagnosis is that AI text is missing things writers don't notice are gone. Research on the structural absences in artificial text identifies losses like embodied authorship and the dialogic appeal to a reader — AI inherits a platform's visibility but never performs the internal appeal to the audience's attention that human writing does, which is exactly the 'aloofness' readers report Does AI writing lack the internal appeal to attention that humans use? Does AI-generated text lose core properties of human writing?. A tool that surfaced these gaps — flagging where a draft has lost the writer's stance or address to a real reader — would convert an invisible loss into a visible editing prompt.

Finally, there's the substrate problem: AI interactions run on context that's mutable, ephemeral, and impossible for users to internalize the way they internalize a stable interface, which argues for treating this as a context-engineering discipline rather than a UI-polish one How does AI context differ from conventional software context?. The thread connecting all of this: unhelpful reliance isn't cured by making AI output better, but by designing for the reader to stay in the loop — making the seams, the distortions, and the missing human signals legible enough that accepting a draft becomes a choice rather than a default.