Scaffold or Cognitive Crutch? A Two-Level Meta-Analysis of AI Scaffolding Effects on Achievement and Self-Regulated Learning in STEM Education

Background. Artificial intelligence (AI) now mediates much of science, technology, engineering, and mathematics (STEM) learning, yet adoption has outpaced theory. The tension is whether AI scaffolding relieves cognitive burden or performs the self-regulatory work autonomy requires, and implementation varies too widely for a single average to describe. Objectives. This review estimates the pooled effect of AI scaffolding on STEM achievement and self-regulated learning, tests whether technological paradigm moderates that effect, and asks whether the literature can evaluate learning after support is withdrawn. Methods. Under PRISMA 2020, four platforms searched in October 2026 (spanning 2021–2026) yielded 21,819 records; 13,786 unique records were dual-screened into an 83-record corpus, then 41 extracted cohorts. Each contributes one harmonised contrast (k = 29; N = 3,709), so within-study dependence is structurally absent. Effects were pooled by two-level random-effects meta-analysis fitted with restricted maximum likelihood and Knapp–Hartung adjustment. Results. The pooled effect was Hedges g = 0.8045 (95% CI [0.5469, 1.0621]), with τ² = 0.4410 and I² = 92.78%. The 95% prediction interval [−0.5846, 2.1936] spans zero, and no moderator reached significance. Notably, 27 of 29 cohorts (93.1%) never withdrew support, and none measured unassisted performance after removal. Conclusions. AI scaffolding yields a large but implementation-dependent gain, consistent with extraneous-load offloading rather than cognitive substitution. The decisive finding is the shape of the evidence base: the field measured a human-AI system in assembly and stayed silent on the competence of the human alone. We name this evidentiary void the Autonomy Gap.