Writing as formation: supervising AI use in early-career medical authorship.

International editorial bodies require authors to disclose the use of large language models in manuscript preparation and to retain accountability for the final text. This guidance is appropriate for formed researchers but does not address what trainees should or should not do during the formative phase of a doctoral thesis, a residency case report, or a first-authored paper. I propose a supervisor-led framework for AI use in medical doctoral and early-career scientific writing, grounded in an argument about professional formation rather than compliance alone; the scope is the writing and interpretation of results, not AI use across all of doctoral research. Writing is one of the settings in which scientific reasoning is exercised and tested, and disclosure is insufficient as a primary control during training because it addresses the reader rather than the writer. When a trainee offloads the first draft of an argument to a model, the formative work that drafting was meant to do does not happen, and the trainee risks becoming a copy-editor of plausible text rather than the author of its argument, a gap no disclosure recovers. The framework rests on three operational moves: a career-stage axis, a sequencing rule, and a revision trace. The career-stage axis makes the same AI operation potentially appropriate for a formed researcher and inappropriate for a first-year trainee. The sequencing rule, applied during training, permits generative AI use only after an unassisted first draft, so that the cognitive work the writing was meant to occasion is performed before the tool touches the text. The revision trace pairs a short disclosure paragraph with an internal log that makes the trainee's tool use inspectable to the supervisor. These moves are organized by a four-tier task typology, related in shape to AI-use rubrics developed in general higher education, delivered as a one-page policy that supervisors or institutions can adapt. The framework is prescriptive in places; the restraint it recommends is formative, and supervisors rather than policy are its primary enforcement mechanism.