AI may make scientists work more, not better, study finds

The paper reframes a widely held belief about AI's promise in knowledge work. Rather than treating time savings as straightforwardly beneficial, the authors apply opportunity cost—a foundational concept in economics—to scientific labor. When a researcher gains an extra hour via AI, that hour is no longer free; it becomes valuable relative to alternative projects. This subtle shift in framing predicts a counterintuitive outcome: more productivity by headcount or submission volume, but less depth per unit of work.

The model isolates this effect by deliberately idealizing LLMs—assuming they introduce no errors and cost nothing—to eliminate confounding factors. This theoretical purity also reveals why the real world shows mixed signals. The OpenAI field report's 60× speedup in coding masked a downstream slowdown in validation and maintenance, a gap the model's framework would predict. The METR study's finding that developers felt faster but worked slower captures the gap between perceived and actual time savings, a disconnect that can still change behavior and policy.

Institutionally, the friction is already manifesting in publishing systems. Fields where LLMs accelerate writing—such as machine learning—are seeing submission surges that overload peer review, and journals are hardening their defenses against AI-generated hallucinations. The paper's call for discipline-specific responses reflects a core insight: acceleration is not uniform. Where AI speeds the wrong phase—early filtering or publishing gates rather than careful analysis—it amplifies volume at the cost of quality, creating what the authors and related software development research frame as a tragedy of the commons, where individual gains impose costs on downstream reviewers and maintainers.