Meta-analysis of 22 studies finds AI in project-based classes helps learners, but gains shrink after bias adjustment

The headline figure is probably inflated. Sensitivity analyses showed small-study effects, and the PEESE correction cut the estimate to g = 0.378, still a positive but much smaller effect.

Product and solution performance, a traditional focus of these teaching methods, was modeled separately and showed a very large pooled estimate of g = 1.958. Its prediction interval crossed zero, so results across settings remain highly uncertain, and the authors call the evidence promising but preliminary.

The authors describe the work as an early-stage synthesis of a nascent field. Their premise is that placing generative AI inside authentic problem solving and meaningful production may build the judgment to evaluate AI outputs and to coordinate AI use across complex workflows.

The study used a three-level meta-analysis to handle multiple effect sizes per study. It pooled four learner-internal outcome domains: achievement and skills, higher-order thinking, affective and motivational outcomes, and collaboration and communication.

Sensitivity checks pointed to small-study effects, meaning the unadjusted g = 0.819 may overstate the benefit. The PEESE specification gave g = 0.378. Exploratory moderator analyses found that estimates varied significantly with participant number and GenAI tool type.

Product and solution performance was analysed separately. Its pooled g of 1.958 may reflect AI-augmented task performance rather than learning, and the prediction interval crossed zero, indicating substantial uncertainty.