Navigating Neurodivergence with AI Chatbots: Benefits, Tensions, and Implications for HCI
Deepak Giri, Erin Brady, Megh Marathe · 2026 · Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems (CHI '26) · doi:10.1145/3772318.3791334
Summary
This CHI 2026 short paper presents a qualitative interview study of how neurodivergent adults use AI chatbots (primarily ChatGPT) in everyday life, and what tensions arise. The authors conducted 23 semi-structured, 50-minute Zoom interviews with participants recruited via Midwestern US university mailing lists and neurodivergence-focused Reddit forums. The sample spanned ADHD, autism spectrum disorder, social anxiety, OCD, learning disorders, sensory processing disorder, and generalized anxiety disorder (many with co-occurring conditions), and skewed young (18-34) and tech-literate. Transcripts were analysed in two phases using a multi-cycle inductive coding pass followed by a deductive top-down pass, iterated across three cycles to produce eleven descriptive codes and nineteen second-level codes, organised under three high-level themes: condition-specific challenges driving AI use, everyday use cases, and tensions with AI. The study positions AI chatbots within broader HCI work on assistive technologies for neurodivergent users, extending prior research on AI for mental health, companionship, communication, and productivity by adding empirical focus on authenticity, privacy, and effects on social relationships. The authors frame their findings in conversation with literature on autistic and ADHD masking, the 'reverse privacy paradox', and concerns that AI-mediated communication can reinforce neurotypical norms.
Key findings
Participants used AI chatbots for condition-specific reasons: emotional regulation (summarising emotionally charged messages to avoid dysregulation), working memory support (summarising long texts and lecture material, especially valuable for ADHD and dyslexia), and self-motivation (breaking overwhelming tasks into steps, reducing procrastination). Common use cases spanned therapy (low-cost, always-available substitute during emotional distress), communication (drafting emails, mediating social conflicts, role-playing a counsellor between parties), school and work (explaining physics word problems, writing user stories), and task planning (evening routines, website-development breakdowns). Several tensions emerged. Twelve of 23 participants described AI as a tool for masking - producing 'neurotypical-sounding' text that hides autistic or ADHD communication patterns - raising authenticity concerns: P8 noted 'if you want to mask, you probably don't want to use something that does sound robotic'. Eight participants used AI as a substitute for human interaction, including to avoid burdening friends, with some describing parasocial relationships with chatbot 'characters'. Participants exhibited the reverse privacy paradox: aware of privacy risks but choosing to disregard them during distress (P2: 'I kind of just pretend I don't see any privacy issues. I just disregard it'). Concerns included AI reinforcing stereotypes of 'robotic' autistic communication, undermining emotional intelligence development through reduced human engagement, and weakening caregiving relationships that normally draw on personal, cultural or religious values.
Relevance
For accessibility practitioners and AI product designers, the paper reframes the neurodivergent AI-chatbot user as someone making urgent, condition-driven trade-offs rather than a frictionless beneficiary. Practical implications include: designing privacy controls that remain visible and usable during emotional distress (since participants knowingly over-share in these moments); avoiding dark-pattern privacy defaults that exploit ADHD hyperfixation; offering authenticity-preserving writing assistance that expands on a user's own voice rather than flattening it into neurotypical templates; and being cautious about AI companion features that may displace valued human caregiving relationships. The paper raises important design-ethics questions about whether supporting masking is the right goal, or whether systems should instead scaffold settings that reduce the need to mask. Limitations are significant: the sample of 23 is young, tech-literate, US-based, mostly ADHD or autistic, and does not include AI non-users or older adults, and the short-paper format limits depth on any single tension. Despite these limits, the work is a useful starting point for accessibility teams integrating LLMs into communication, mental-health support, or productivity tools.
Tags: neurodivergence · AI chatbots · ADHD · autism · masking · working memory · emotional regulation · large language models · mental health · privacy · authenticity