Supervision and doctoral education

Each claim cites its evidence; a second model has checked it against that evidence.

1. 01

   Across two independent studies of doctoral supervision, hidden GenAI use emerges as the central relational problem: Boyd & Harding 2025 find students "(often covertly) using GenAI tools to support their emotional and discipline specific development," while Lai et al. 2025 trace "conflicting expectations regarding supervisory responsibilities among students, supervisors and peers"—and both conclude that supervision must become transparent and collaborative rather than prohibitive.

   ◐Partly supported3 sources · 2 papersTrace →
2. 02

   Han & Liu 2026 provide survey evidence that supervisory support "weakened the negative associations between AI dependence and research autonomy and innovative behaviour," a micro-level, empirical analogue of Park et al. 2026's model result that mandatory practice preserves capability—together suggesting structured human oversight is the practical lever against AI-induced deskilling.

   ✓Supported2 sources · 2 papersTrace →
3. 03

   Gao & Zhang 2026 specify what the "epistemic supervision" called for by Han & Liu 2026 looks like in practice: supervisors had students "mark AI-influenced parts and explain why suggestions were accepted or rejected," which "pushed [them] to connect AI outputs with source evidence, methodological assumptions, and argumentative responsibility."

   ✓Supported2 sources · 2 papersTrace →
4. 04

   Delegation buys immediate performance at the cost of human engagement: in Hao et al. 2026 the Delegated Reasoning group "achieved the highest task performance" yet the Concerted Interpretation group "reported significantly greater use of self-regulation strategies," and Wu et al. 2025 show the performance boost "did not persist in subsequent tasks performed independently by humans" while intrinsic motivation fell—a small-scale signature of Park et al. 2026's "enrichment paradox."

   ◐Partly supported2 sources · 2 papersTrace →
5. 05

   Field and laboratory evidence conflict on whether AI assistance builds durable human skill: Brynjolfsson et al. 2024 "find evidence that AI assistance facilitates worker learning," with the largest gains for less-experienced, lower-skilled workers, whereas Wu et al. 2025 find gains vanish in solo work and boredom rises—so the conditions under which AI collaboration converts into lasting human learning remain unidentified.

   ✓Supported2 sources · 2 papersTrace →
6. 06

   Two recent scale-development papers converge that legacy instruments miss what matters in AI collaboration: Dai et al. 2026 "redefine student agency as a process of human-AI coagency grounded in adaptability, epistemic responsibility, and distributed inquiry," and Sidra & Mason 2025 show "Collaborative AI Metacognition explained significant variance beyond that explained by general Metacognition."

   ✓Supported2 sources · 2 papersTrace →
7. 07

   Park et al. 2026 argue that deskilling belongs on the top-tier AI-governance agenda alongside alignment, bias, and misuse: their model "reframes the policy question from 'should we adopt AI?' to 'how capable should AI be allowed to become in specific domains,'" warning that "incremental AI improvement can trigger discontinuous societal consequences" near a critical threshold K*≈0.85.

   ✓Supported2 sources · 1 papersTrace →
8. 08

   Capability-threshold governance of the kind Park et al. 2026 propose presumes reliable measurement of what AI systems can do, but Mullens & Shen 2026 show that foundation is shaky: "half of the hardest GPQA items contain chemically or logically indefensible answers," and "answering the persona question requires evaluation infrastructure the field does not yet possess."

   ✓Supported2 sources · 2 papersTrace →
9. 09

   The emotional reliance on GenAI that Boyd & Harding 2025 document in doctoral supervision—"the emotional support ascribed by students to the technology"—is independently flagged by Zhu et al. 2026 as an unmet design challenge, since "deployment in emotionally sensitive contexts" is among the agent capabilities "not fully captured by existing principles."

   ◐Partly supported2 sources · 2 papersTrace →
10. 10

    As AI spreads across professions, the scarce human skill shifts from operating one tool to orchestrating many: Calboreanu 2026 centers "the practitioner's ability to select, configure, and compose tools across domain boundaries," with prompt and context engineering transferring "as portable cross-tool metaskills"—echoing Bankins et al. 2024's conclusion that "AI literacy and other human-centered skills will play an increasingly important role in future workplaces."

    ✓Supported3 sources · 2 papersTrace →