The Cost of Not Adopting: A Priority-Race Model of Generative AI in Research Teams, with a Synthesis of Benefits and Risks

LLM assistants such as ChatGPT and Claude have become routine research tools: in 2025, 84% of surveyed researchers reported using AI tools, up from 57% a year earlier. Experiments document substantial but task-dependent time savings, and bibliometric studies associate adoption with higher output, although that association is disputed. We ask what a research team loses by not adopting when its competitors do. Building on classical patent-race models, a stylized priority-race model combines task-level time savings, the jagged frontier of AI competence, checking costs and undetected errors. It yields a closed-form win probability, a break-even condition for adoption, and a √k law by which multi-stage, winner-take-all competition amplifies modest speed advantages. In scenarios informed by published effect sizes, moderate and strong adopters win two-team races with probabilities 0.63 and 0.82, complete 17% and 55% more projects over time, and meet a deadline that a non-adopter meets half the time with probabilities 0.68 and 0.92. When losers of a race keep most of their credit, as citation evidence on scooped papers suggests, the per-race credit advantage narrows to 6–16%, and the cost of abstention operates mainly through throughput and deadlines. Abstention is rational when AI slows experts down or checking absorbs the savings, and the value of verification rises with the risk of misapplying AI. Within this framework, LLM assistants become competitively indispensable in time-sensitive research where rivals obtain reliable net gains. Calibrated, verified and disclosed use is the strategy the evidence supports.