AI Tutors vs. Human Teachers A Systematic Review of Learning Outcomes (2020–2026)

Can an artificial intelligence tutor teach as well as a human teacher, and under what conditions does it help or harm learning? This review synthesises evidence from 2020 to 2026 on AI tutors, human teaching, and combinations of the two, drawing on randomised trials, quasi-experiments, meta-analyses, systematic reviews, deployment evaluations, and policy and survey evidence. AI tutors can produce real learning gains, but the gains depend heavily on design and on human presence. A purpose-built tutor that scaffolds step by step, withholds direct answers, and operates inside a classroom or alongside a tutor outperformed, in one trial, an active-learning class. A general chatbot handed to students without guardrails does the opposite: practice scores rise while the tool is open, then unassisted exam performance falls below that of students who never had access. The strongest moderator is teacher support. In one meta-analysis, generative AI interventions with teacher support averaged g = 1.426, against g = 0.077 without it. Engagement rises with AI use, but engagement and learning repeatedly come apart: students like the tool, produce better work with it open, and still retain less weeks later. Equity results cut both ways. AI tutoring has produced strong gains in low-resource settings, yet platform use in one large observational study amplified existing advantages tied to affluence and urban location. The useful question is not whether AI replaces teachers. It is which teacher, guardrail, and assessment arrangements turn an AI tutor from a crutch into a tutor.