Artificial Intelligence in Higher Education: Transformations, Empirical Evidence and Challenges for Student-Centred University Learning

The rapid evolution and integration of Artificial Intelligence (AI) technologies, particularly following the advent of Large Language Models (LLMs) and Generative AI (GenAI), have profoundly transformed teaching, learning and management practices in higher education worldwide. This paper aims to synthesise the empirical evidence available between 2020 and 2026 regarding the impact of AI in higher education, examining the prevalent technologies, pedagogical mediation, levels of human–AI collaboration, adoption patterns and ethical-normative implications. A systematic literature review was conducted in rigorous accordance with the PRISMA 2020 Statement. Six international scientific databases were consulted: Scopus, Web of Science, ERIC, SciELO, PubMed and IEEE Xplore, identifying an initial corpus of 1,247 records. After the screening process for duplicates, title/abstract and full-text reading, 45 primary studies focused on university educational impact were selected, analytically complemented by global bibliometric and meta-systematic reviews of the period. Methodological quality was assessed using the MMAT (2018) tool and the certainty of the evidence via the GRADE approach. It is concluded that the predominant AI technologies in the literature are large language models (75.6 %), intelligent tutoring systems (31.1 %) and learning analytics tools (24.4 %). The principal applications concentrate on personalised learning (53.3 %), student motivation and engagement (46.7 %) and academic performance (42.2 %). However, evaluation of the autonomy continuum reveals a severe implementation gap: 45.56 % of integrations correspond to Substitution modalities (where AI replaces teaching functions), 39.05 % to Augmentation, 10.65 % to Toolification and only 4.73 % achieve genuine Human–AI Synergy. Longitudinally, student adoption reached near-universal levels (95 % in 2026), although gaps in ethical literacy and tensions surrounding academic integrity persist. The methodological quality of the corpus was high in 62.2 % of the studies, yet the global GRADE certainty was classified as low owing to the predominance of cross-sectional designs and heterogeneity of metrics. It may further be concluded that AI possesses a proven transformative potential to optimise university learning, yet its effectiveness does not reside in mere technological adoption, but rather in intentional pedagogical mediation (the PMAISE model), the evolution of teaching practice towards the AI-TPACK framework and the establishment of executable governance policies (Policy-as-Code). A human-centred integration framework based on four pillars and a five-phase operational strategy with a detailed matrix of success indicators for institutional scalability is proposed.