Build expertise first: why PhD training must sequence AI use after foundational skill development

Generative AI tools have arrived in PhD training environments faster than principled frameworks for their use. The debate has polarized between enthusiasts who argue trainees must adopt AI immediately and critics who warn of fundamental damage to learning. Both miss the key question: not whether trainees should use AI, but when. The answer requires distinguishing two fundamentally different kinds of automation. Previous technologies like calculators, statistical software, and search engines automated mechanical execution while preserving the cognitive work that constitutes learning. Generative AI is categorically different: it automates reasoning, synthesis, and judgment themselves. This distinction matters enormously for training, because PhD education is not primarily about task completion but about developing capacity for independent scientific thought through sustained cognitive struggle. This creates what we term the verification paradox: trainees cannot meaningfully verify AI outputs because verification requires the domain expertise they are still developing. Using AI before that expertise exists bypasses the developmental process that builds it while producing polished outputs that mask the gap. The solution is sequencing. Trainees should build foundational expertise through deliberate, feedback-driven practice before using AI to augment it. The threshold is demonstrated independent mastery: the ability to complete tasks, explain reasoning, and catch errors without assistance. Once crossed, AI use becomes not just acceptable but genuinely productive. PhD programs, mentors, and professional societies urgently need community standards built on this developmental framework rather than ad hoc policies shaped by convenience. An implementation of this article's conceptual framework as a detailed guide with task-specific protocols is available at https://doi.org/10.5281/zenodo.18452319.