Answers are grounded in this paper and cite their sources.
Introduction Intelligent tutoring systems have been synthesized repeatedly, yet estimates of their effect on academic achievement continue to diverge. This systematic review examined whether such systems improve academic achievement relative to instruction without them. Methods We included 89 studies published between 2015 and 2026, identified through Scopus, Web of Science, and ERIC. A system qualified only when it adapted instruction from performance data collected during use, and a study was eligible only when it compared the system with an arm that did not use it and reported a between-group test on an achievement measure. Studies in which every arm used an adapting system addressed which design features improved learning, not whether such systems improved achievement; 119 such studies formed a design-feature group that lay outside the synthesis and is listed separately. Direction of effect and statistical certainty were coded as independent fields, the comparison condition was coded from the arm actually contrasted, and every coded value was verified against the full text. Sensitivity analyses examined alternative treatments of the coding rules, including the effect of counting 15 design-feature studies, whose comparison arm was another version of the same software, as active comparisons. Results The point estimate favored the system in 83 of the 89 studies (0.93; 95% CI 0.86-0.97). A significant advantage was reported in 71 studies (80%), there was no significant difference in 12, and there was a mixed result in 6; none favored the comparison condition. The proportion lay between 72% and 80% under each alternative treatment of the coding rules. Seventy-three studies compared the system with conventional instruction or no intervention and 13 with an alternative technological or structured intervention; 79% and 85%, respectively, reported an advantage. Had the 15 design-feature studies been counted as active comparisons, the latter proportion would have fallen to 39%. Achievement was measured with a researcher-developed instrument in 63% of studies and a standardized instrument in 18%, and a delayed measure was administered in 12. Positive effects were reported in 58% of single-session studies and 89% of those lasting 13 weeks or longer. All 21 studies of systems built on large language models were published in 2025 and 2026. Discussion The direction of effect consistently favored the system, and no study favored the comparison condition. The drop to 39% under the alternative coding reflects the boundary of the intervention rather than differences in instruction. The higher proportion of positive effects in longer studies is inconsistent with a novelty account.
No takes yet. Share an insight, caveat, or question.
Игоревна et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: