We Ran a Research Sprint With AI. The AI Worked. The Research Didn't

arXiv:2611.03946v1 Announce Type: new Abstract: Generative AI tools for research are widely deployed, yet individual productivity gains do not guarantee better team research. We argue that how AI use is structured matters alongside access. Personal AI tools may amplify two failures under deadline pressure: a consumption bottleneck, in which generation outpaces team-visible evaluation, and AI-induced research silos, in which private human-AI loops displace shared decision-making. We motivate this position with a mechanism-generating pilot: a 100-minute research sprint with twelve researchers in three independently operating teams, within a seventeen-person session. Both patterns recurred across teams and evidence sources.