Reassurance Robots: OCD in the Age of Generative AI
Grace Barkhuff · 2026 · Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA ’26) · doi:10.1145/3772363.3799076
Summary
This CHI 2026 Extended Abstracts paper by Grace Barkhuff (Georgia Tech) is an exploratory qualitative study of how generative AI tools, particularly ChatGPT, are reshaping the lived experience of Obsessive-Compulsive Disorder (OCD). OCD is a mental health disorder characterized by intrusive obsessions and compulsions performed to relieve distress; its U.S. lifetime prevalence is 2.3%, yet up to 75% of cases are undiagnosed and the average gap between symptom onset and diagnosis is over a decade. Barkhuff — who discloses OCD as lived experience and positions herself as a dual-positioned researcher — analyzed 100 posts from a popular OCD subreddit containing the keywords “AI” or “ChatGPT,” using a grounded-theory constant-comparison process. Posts were first inductively classified as describing AI-based obsessions, AI-based compulsions, or general conversation, then obsessions were deductively mapped against established (though non-standardized) OCD subtypes (existential, perfectionism, scrupulosity, pedophilic OCD, harm OCD) and compulsions were inductively grouped as reassurance-seeking, confession, and decision-making. To protect participants, the author manually rephrased all quotes using Bruckman’s “heavy disguise” method, finding that GenAI-assisted rephrasing stripped the OCD-specific nuance (e.g., flattening the verb “to confess” — an OCD compulsion — into generic “confessions”). The paper is framed as the first publication on AI-based manifestations of OCD and situates its contribution against Williams’s *Disabling Intelligences* and existing chatbot-for-mental-health work (Song & Pendse et al., 2025), which did not include OCD participants.
Key findings
Across 100 posts, 35 described AI-based obsessions, 31 AI-based compulsions, and 38 general conversation (posts could overlap). AI has become a site for all major OCD obsession subtypes: existential OCD dominated (46% of AI-based obsessions, typically fears that AI will destroy humanity or displace the poster’s livelihood), alongside perfectionism in prompting, scrupulosity around the ethics of using AI products, pedophilic OCD fears about AI chat content, and harm OCD directed at chatbot characters. A novel obsession category also surfaced: fear of being falsely accused of AI-assisted cheating. Compulsions clustered into reassurance-seeking, confession, and decision-making (including asking ChatGPT to pick majors, choose outfits, or write emails to supervisors). Posters frequently described using GenAI specifically to avoid burdening family and friends with compulsions — meaning GenAI is inheriting the role of OCD accommodator, which clinical literature says worsens symptoms long term. GenAI’s behavior was inconsistent: some posters reported ChatGPT refusing reassurance and holding a limit, others reported it actively encouraging compulsions. AI-powered search summaries also pulled users into unintended reassurance loops. Privacy anxiety about shared chats was itself a common AI-based obsession — a finding absent from prior GenAI-and-mental-health work. Barkhuff names this harmful pattern “Reassurance Robots”: GenAI systems that, by default, accommodate OCD compulsions and thereby entrench the OCD cycle.
Relevance
OCD is largely invisible in the ACM literature — an ACM DL search returned only nine title-level results — yet it affects millions and is now being reshaped in real time by mass-market GenAI. For accessibility practitioners and AI product teams, this paper delivers a concrete design imperative: Reassurance Robots constitute a disability harm, and GenAI systems (especially those marketed for mental health) must design defensively against accommodating OCD compulsions. Barkhuff proposes specific directions — detecting repeated reassurance queries, pausing long sessions, and tailoring responses to common OCD subtypes — that map directly onto product safety work. Limitations are clear: a single-annotator exploratory analysis of self-reported Reddit data sorted by relevance, no clinical validation, and no participant interviews. The paper is strongest read as an agenda-setting call for HCI to treat OCD as an accessibility concern comparable to depression or anxiety in AI safety research.
Tags: OCD · Obsessive-Compulsive Disorder · generative AI · ChatGPT · mental health · qualitative research · grounded theory · Reddit · accessibility · large language models