## A5. HITL AI Scaffolding — Autonomy-Supportive Design **Core finding:** AI scaffolding must support autonomy, not replace executive function. Controlling or fully automated approaches undermine the self-regulation skills they aim to support. The distinction is not philosophical but empirical. **Self-Determination Theory (SDT) framework for ADHD:** - **Champ, Adamou & Tolchard 2022** (*Psychological Review*): Proposed a complete SDT-based framework for ADHD, arguing that autonomy, competence, and relatedness needs explain self-regulation patterns better than deficit models. - **Champ et al. 2025** (*JMIR Formative Research*): ADAPT randomised feasibility study with **20 adults from an NHS ADHD clinic**. **91.6% intervention completion**. Clinically significant improvement in psychological distress (p = .01) and significant ADHD symptom reduction (p ≤ .01). Demonstrates that autonomy-supportive scaffolding works in clinical practice. **Critical review of existing ADHD tools:** - **Spiel et al. 2022** (*ACM CHI '22*): Most ADHD technology is "shaped by research aims which privilege neuro-normative outcomes." Time-management interventions frequently cause stress and frustration. Participatory design with ADHD individuals leads to **fundamentally different design outcomes** (e.g. conceiving time as "stretches of activities" rather than clock-based units). Explicitly documents harm caused by surveillance-like monitoring and intrusive alarms. - **Carik et al. 2025** (*ACM GROUP '25*): LLM use across **61 neurodivergent Reddit communities**, identifying 20 use cases. ADHD users primarily sought help with organisation, planning, and prioritising. LLM responses are frequently **"overly neurotypical"** and not calibrated for neurodivergent cognition. Users expressed significant concern about overreliance. **Longitudinal case evidence:** - **Mittler 2025:** 42-year-old neurodivergent student with severe executive dysfunction. Over 4 semesters using strategically integrated AI tools, GPA rose from **1.85 to 3.35**. Psychological trajectory shifted from anxiety to sophisticated "process awareness" — the student internalised external scaffolds. - **Azevedo et al. 2022** (*Frontiers in Psychology*): Decade-long MetaTutor programme, 100+ college students. **Adaptive pedagogical agents that prompt metacognitive strategies** (rather than completing tasks) produced significantly better learning outcomes. **Five design principles from the literature:** 1. **Scaffold, don't automate** — prompt metacognitive strategies rather than completing tasks for the user 2. **Co-regulate, don't correct** — nudges should be reflective ("What were you working on?") rather than directive ("You should be working on X") 3. **Adapt to fluctuating states** — detect attention shifts and adjust support intensity dynamically 4. **Keep the human in the loop** — every AI suggestion requires user confirmation, building executive function rather than atrophying it 5. **Design with, not for** — participatory design with neurodivergent users produces fundamentally different and better outcomes **Implication for Magnotia:** The AI agent must be visible, conversational, and interactive — but must never override user autonomy. Every suggestion requires confirmation. The human-in-the-loop feedback mechanism builds metacognitive awareness over time. Users should eventually internalise Magnotia's scaffolding patterns and need them less — that's a feature, not a failure. LLM prompts must be calibrated for neurodivergent cognition, not neurotypical assumptions.