JMIR Publications

ABSTRACT

Research supervisors face an unresolved question: how can trainees use generative language models, which already draft near-publication-quality manuscripts, abstracts, grant proposals, and slide presentations, without bypassing the cognitive struggle on which research competence is built? This viewpoint argues that the protective factor for cognitive retention and engagement is AI deliberately structured to demand thinking before it will assist, and it describes a working coaching architecture built on that principle, together with the literature that motivated it. Three custom AI agents were developed, one for each of three output categories (abstracts and manuscripts, slide presentations and posters, and grant and scholarship proposals), all sharing one architecture: intake calibrated to the target journal or funding agency; adversarial review by two to four expert reviewer personas selected based on document type; Socratic tutoring that withholds corrected text; iterative re-review with explicit tracking of resolved and unresolved issues; and a structured reflective debrief before any document reaches the supervisor. The design rests on established learning science, chiefly cognitive apprenticeship, so that struggle precedes assistance rather than being replaced by it. Of 20 trainees in this interprofessional laboratory (undergraduate and graduate students, postdoctoral fellows, clinical fellows, and surgical residents), 10 have used the agents to date. Uptake was favorable once initial unease about exposing early-draft errors to an automated reviewer gave way to relief at correcting them privately, before the supervisor ever saw them. The most unanticipated outcome was a new form of reflexive dialogue: trainees increasingly returned with counterarguments, corrections to the coach’s own critique, and ideas the supervisor had not previously considered, an effect the laboratory now plans to measure prospectively. This pattern is consistent with recent evidence that partnership-oriented engagement with AI increases critical vigilance and strategic delegation simultaneously, with both predicting transformative learning rather than trading off against one another. Because the agents make explicit the reviewer reasoning, ordinarily transmitted only through proximity to senior mentors, the architecture also offers a path toward widening access to research mentorship for trainees at low- and middle-income-country partner sites. The viewpoint offers other academic supervisors a transferable architecture.