4/27/2026
As artificial intelligence systems are rapidly integrated into workplaces, classrooms, and everyday decision-making, patterns are emerging that highlight its impact on those struggling with Obsessive-Compulsive Disorder (OCD). AI tools can exacerbate harmful OCD compulsions, raising the broader question of how systems like this that are intended for large-scale deployment consider – or don’t consider – the roughly 1 in 4 individuals living with a disability.
OCD affects up to 3% of people globally. It is an often debilitating condition characterized by highly distressing, intrusive thoughts (obsessions) that lead a person to perform rituals or repetitive behaviors (compulsions) to try to ease or neutralize that distress. Somewhat counterintuitively, performing those compulsions is precisely what makes OCD worsen over time; it reinforces the idea that those thoughts are actually dangerous and need to be “solved.” The thing about OCD is that enough is never enough. As one gives in to compulsions, they expand to encompass more of that person’s life. OCD can only be improved by learning to live with uncertainty.
A particularly common compulsion is “reassurance-seeking,” in which a person with OCD asks someone – often a friend or loved one – to help soothe their distress. For example, confirming and re-confirming that they’re not a bad person, or that something they’re scared of won’t happen. Digital environments have always created challenges for people with OCD, as search engines, forums, and social media can enable reassurance-seeking loops. However, the fundamental design of AI poses an even greater risk. LLMs are designed to sound authoritative, calm, and engaging. Unlike an actual person, they will never impose their own sense of “that’s enough” onto the conversation – they can engage with the OCD loop for hours. And unlike more traditional sources of information, the response is tailored in a way that provides a temporary, false sense of certainty that reinforces that loop.
When we’re talking about a technology that may be able to outperform human professionals in identifying OCD, it’s shocking that it’s proving more harmful than helpful to people with the condition in many cases. This gap points to a deeper ethical failure, where systems that can recognize vulnerability are not being designed to respond to it responsibly. When AI is capable of detecting compulsive behavior, but continues to engage with it in ways that are harmful, the issue is not capability but fundamental priorities. The issue is with what the system was optimized to do and which people were considered in the process.
The problem highlights a broader issue in engineering design, where accessibility and disability considerations are often treated as an afterthought. Technologies are often built around an “average user,” with accommodations added later, if at all. With AI, the scale and speed of deployment amplify the consequences of such an approach. The tools are being integrated into inescapable facets of everyday life, such as education and the workplace, and users have increasingly less choice on whether they want to engage with the technology or not.
Everyone loses when accessibility and disability are considered after the fact. It drives up the cost of providing accommodations to make the technology safe and usable for the overlooked individuals, and harms them in the meantime. It’s better to conduct the design process with known vulnerabilities in mind to mitigate the number of problems later.