When ChatGPT Is Down, So Are Our Brains: Cognitive Fragility Following AI Disruption in Knowledge Work

Larry Zhiming Xu, Gabriel Velez, Jacklynn Fitzgerald, Jungmin Lee, Terence T. Ow

As generative AI becomes embedded in everyday knowledge work, it increasingly functions not merely as a tool but as cognitive infrastructure—something users expect to be continuously available, responsive, and reliable. While recent discussions of AI-related “brain rot” have focused on gradual cognitive decline through long-term dependence, less is known about what happens when expected AI support suddenly fails. This study introduces the concept of cognitive fragility to describe a short-term vulnerability in performance and judgment that may emerge when individuals who anticipate AI assistance must suddenly work without it. Drawing on theories of cognitive offloading, technology dependence, and expectation violation, we argue that AI disruption does not simply return users to an unaided baseline. Instead, it may destabilize task strategies organized around anticipated automation support, producing immediate performance costs and increasing susceptibility to later automation errors. We test this argument through a pre-registered between-subjects lab experiment with undergraduate participants from a Midwestern U.S. university (N = 71). Participants were randomly assigned to one of three conditions: no AI access, continuous ChatGPT access, or promised ChatGPT access that was unexpectedly disrupted due to a simulated IT issue. Participants completed timed verbal reasoning, quantitative reasoning, and writing tasks using GRE-based materials. Afterward, they were offered a chance to improve their performance using a purported “advanced AI model” that, unknown to them, generated incorrect answers. Their willingness to overwrite their original responses with these faulty outputs served as a behavioral measure of automation error susceptibility. Preliminary results show that continuous AI access significantly improved quantitative performance, while verbal task performance did not differ across conditions. However, when AI access was expected but disrupted, participants produced significantly less writing and took longer to complete the tasks compared with those who never expected AI support. More importantly, participants in the AI-disrupted condition were more likely to accept erroneous AI-generated answers afterward, suggesting that disruption heightened uncritical reliance rather than increasing vigilance. Together, the findings indicate that AI reliability failures may create downstream cognitive and judgment costs beyond the immediate loss of assistance. The study contributes to research on cognitive offloading, automation dependence, and AI-augmented work by shifting attention from the effects of AI use to the consequences of AI withdrawal. It also highlights the organizational need for fallback routines, resilience training, and accountability mechanisms when AI systems fail.