Learn First, Then Use AI: Testing an Expert Consensus on Generative AI and Learning Against the Empirical Evidence

Debate about generative AI in education often asks whether it is good or bad for learning. A 2026 expert consensus report from the University of Minnesota, UNC-Chapel Hill and UC Irvine argues that this question is too blunt. Thirty experts used a preregistered Delphi method to rate 65 learning processes for how likely general-purpose generative AI is to enhance or disrupt them, reaching consensus on 42. The pattern they describe can be read as a sequence: AI threatens the early work of taking in new information, can support the consolidation of what has been learned, and does not yet reliably support higher-order thinking and transfer. The report also rates the strength of the existing evidence as low. This review asks whether the experimental and observational studies published so far are consistent with that consensus. It maps a field experiment with high-school mathematics students, a randomised con‐ trolled trial of an AI physics tutor, an EEG study of essay writing, and survey research on cognitive offloading onto the three stages of the sequence. The studies largely agree with the experts. Unrestricted answer-giving harms acquisition. Tutor-style designs built on learning science can strengthen consolidation. Offloading thinking is linked with weaker critical thinking. We summarise this as a 'learn first, then use AI' principle, link it to established learning science, and set out implications for teachers, school leaders and researchers.