Answers are grounded in this paper and cite their sources.
Generative artificial intelligence (GenAI) can improve what university students produce without showing what they can do without it. This systematic review examined when GenAI improves task performance, when that improvement becomes learning that students can demonstrate independently, and how such learning can be assessed. We searched OpenAlex for studies published between November 2022 and October 2026, supplemented by citation chaining, and included 411 controlled comparative studies and 99 systematic reviews (511 reports). Screening, data extraction and risk-of-bias appraisal were carried out by large-language-model reviewers under the author's supervision, without human verification. Each outcome was coded by whether students could use AI when it was measured and by outcome state (AI-assisted performance, immediate independent performance, delayed retention or transfer), and the studies were organised in an evidence-and-gap map by the role the AI played. In 292 of the 411 primary studies (71%), no outcome allowed us to establish whether AI was available at measurement, so only 92 studies entered the map. Most mapped evidence concerned performance with AI available (59 studies) or immediate independent performance (32 studies); only 4 studies measured delayed retention and 5 transfer, and no mapped study was at low risk of bias. Within single studies, gains obtained with AI available often did not carry over to tests taken without AI, whereas several contrasts on delayed or transfer outcomes favoured designs that required an independent attempt or self-evaluation before AI use. None of six recent meta-analyses examined in the introduction reported whether outcomes were measured with or without AI. Gains measured with AI available should be read as evidence of performance, not of learning. We propose productive cognitive friction, intellectual ownership and dual assessment as hypotheses for future trials, and call on primary studies to report AI availability and timing for every outcome. Published in Journal of AI-Generated Reviews as article JAIGR.2026.0001, version 1. The review was generated by AI (Claude Opus 5.5) and assessed by AI with a human editor's decision; it was not peer-reviewed. It is not licensed for reuse, and its file here is not public; read it at the address above.
No takes yet. Share an insight, caveat, or question.
Han Ding (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: