The rapid development of generative artificial intelligence is reshaping the education ecosystem, but there is a disconnect between the leap in technological capabilities and the release of educational value, despite 85% of schools introducing generative AI tools and only 23% of cases realizing significant learning gains. By constructing a meta-analytic framework of multi-dimensional moderating effects, this study systematically analyzed data from 80 empirical studies, and used the Hedges' g effect size standardization method to reveal the differences in the efficacy of generative AI in different instructional environments, subject categories, learning styles, and interaction designs. The results show that the overall educational effect size of generative AI is Hedges' g = 0.58, which is significantly higher than that of traditional tools (Δg = 0.23, p < 0.001), especially in multimodal interaction and collaborative learning scenarios. However, the effectiveness of the technology depends on strict conditions, including model parameters exceeding 10 billion (e.g., GPT-4) and the amount of feedback information ≥ 50 words; otherwise, the effectiveness may be significantly degraded. The study suggests building a dynamic tracking system to cope with technology iteration and deepening brain mechanism research to optimize instructional design. In the future, educational practices need to unleash the disruptive potential of generative AI in the resonance of technological reliability, cognitive adaptability, and teaching systemicity, so as to promote education from efficiency improvement to paradigm innovation.