Making AI Annoying on Purpose: When Helpful Tools Don't Always Help

Synthesis: Konradt, Boote, and Taub (2026) report a design-based study in which four high-school students used a constrained AI writing system of teacher-delimited chatbots designed to ask questions rather than generate text, oriented through the TRACE model (Target, Refine, Assess, Cycle, Extend) over a nine-week argumentative-writing unit. Students evolved from passive AI consumers to strategic evaluators who pushed back on AI outputs, developed 'prompt literacy,' and strengthened counterargument development and evidence integration, suggesting that strategic constraint — not unrestricted generativity — was associated with greater self-AI Regulation in Education and sustained engagement.

Key Findings

• A constrained AI system using teacher-delimited, question-asking chatbots promoted 'productive friction' that required students to defend and refine their thinking.
• Over nine weeks students shifted from passive AI consumers to strategic evaluators who actively pushed back on AI outputs.
• Students developed 'prompt literacy' — crafting targeted requests and critically evaluating responses — while strengthening counterargument development and evidence integration.
• Constraint, rather than unrestricted generativity, was associated with increased self-regulation and sustained engagement with argumentative contexts.

What this means for practice

• Teachers. Deploy question-asking, teacher-delimited chatbots that refuse to generate student prose, keeping learners inside the effortful parts of argumentative composing.
• Instructional designers. Sequence constraints progressively across a unit rather than granting unrestricted GenAI access, so the tool creates productive friction instead of completing students' thinking.
• Teachers. Protect time for weekly reflections and targeted mini-lessons, the mechanisms through which students turned pushback on AI output into counterargument and evidence-integration gains.
• Researchers. Pair interaction logs with writing samples and interviews when studying constraint-based design, since the evidence here is qualitative and process-oriented rather than outcome measurement.

Limitations

• Four volunteer high school students, chosen by maximum variation sampling on baseline writing ability and AI attitudes, in two sections of a single AP Capstone course.
• A nine-week design-based study with no control group, so changes in argumentative reasoning cannot be causally attributed to the constrained system.
• Evidence rests on self-report and researcher-coded data — weekly reflections, interviews, writing samples, and AI interaction logs — analyzed through thematic analysis.

Connected Concepts

• Writing
• Self-Regulated Learning
• Generative AI
• Scaffolding
• Learning Design
• Desirable Difficulties

Connected Articles

• [making-ai-tutoring-productive-mastery-math-2026] — designing AI tutoring for productive mastery in math
• [ai-making-us-stupid] — critiques of AI's effect on cognition

Citation

Konradt, L., Boote, D. N., & Taub, M. (2026). Making AI annoying on purpose: when helpful tools don't always help. Computers and Education Open, 10, 100343.