Stop Writing for Me: Generative Refusal in AI Tools for Thought

Synthesis: Position paper exploring "Generative Refusal" — AI tools that strategically withhold text generation to demand user articulation, functioning as a Maieutic Partner rather than a cognitive offloading tool. Argues that in domains where the labor of articulation is central to craft, AI should enhance rather than bypass human cognition.

Key Findings

• The paper critiques current GenAI design paradigms that prioritize cognitive offloading — generating text on the user's behalf — which risks eroding the constructive thought process essential to artistic training.
• It proposes Generative Refusal: strategically withholding text generation to demand user articulation, positioning AI as a Maieutic Partner that returns cognitive friction to the user and shifts the interaction from delegation to active articulation.
• The framework is instantiated in Actor's Note, a character-journaling tool for actor training that generates context-aware questions instead of draft text.
• A field study of Actor's Note suggests the constraint significantly reduced cognitive burden while fostering a residual effect of internalized questioning habits.
• The motivating context is character journaling — a foundational tool for thought in actor training that is valuable but hard to sustain: actors frequently abandon it not from lack of material but because the high cognitive load of initiating reflection after exhausting rehearsals creates a barrier, often manifesting as the "blank page" problem.

The Maieutic Interaction Framework

Instead of bypassing cognition, the Maieutic Interaction Framework returns cognitive friction to the user: the system withholds drafts and generates questions that prompt reflection, inverting the efficiency-oriented default in which generative tools write for users. For Creativity support and Writing, the framework suggests that the absence of generated text can be a deliberate design feature rather than a deficiency, protecting the constructive thought process that the labor of articulation is meant to build.

What this means for practice

• Designers. Treat refusal as a designed mechanic rather than a missing feature: wherever the labor of articulation is the point of the activity, have the tool return scaffolded questions instead of drafts, as Actor's Note does with three rehearsal-stage-specific questions per entry, so the constructive thought process AI would otherwise bypass is protected against generative efficiency.
• Instructors. Evaluate such tools by internalization rather than output speed — after assistance was removed, participants recalled or self-generated the AI's questioning style at M = 4.87 on a 7-point scale — and build a post-tool performance check into any AI-supported learning task.
• Edtech designers. Match friction to the workflow phase as the 14-day crossover deployment found: low-friction prompts to break early inertia, harder challenging questions later to counter fixation, since the early-AI group showed a momentum effect while the late-AI group gained Narrative Transportation (p = .0128).
• Instructors. Frame AI reflection spaces as non-judgmental sandboxes so students disclose unpolished thinking, then route that material back into human feedback loops instead of letting private AI work replace them.

Limitations

• The empirical core is one 14-day in-the-wild study with 29 professional and student actors in a single discipline (actor training), and it compares AI-assisted with unassisted journaling rather than with a text-generating tool.
• Primary outcomes are self-reported daily surveys of cognitive burden, intrinsic motivation, and acting confidence plus post-study interviews, so the reported effects rest on subjective measures.
• The internalization finding (M = 4.87 of 7) is a single post-study self-report with no delayed follow-up, and the log measures (lexical diversity, emotion words) describe written output rather than assessed learning.
• Extension beyond theater is argued by analogy: the paper raises coding and academic writing as workshop discussion questions rather than testing refusal designs there, and its four design implications are proposals from a 4-page position paper rather than validated interventions.

Citation

Sora Kang (2026). Stop Writing for Me: Generative Refusal in AI Tools for Thought.