The Cognitive Bottleneck in AI-Accelerated Research

AI tools are supposed to make research faster. In my experience, they do, but only for certain tasks. Over six weeks, I tracked 243 AI-assisted sessions in my computational biology work. The pattern that emerged surprised me: I could draft emails and generate code dramatically faster, but when it came to actually understanding complex results, the speedup largely vanished. This paper documents that observation and tries to make sense of it. Drawing on cognitive load theory and what we know about human information processing limits, I argue that we may be hitting a “cognitive bottleneck,” a point where AI output exceeds our capacity to meaningfully process it. I propose a framework called PACE for managing this bottleneck, though I should be clear: it’s untested, and this whole paper is based on one person’s experience. Take it as a hypothesis worth investigating, not a conclusion.