Use quantity or use quality? AI engagement depth, academic performance and the Matthew effect in undergraduate capstone projects: Evidence from 875 graduating students

Abstract Generative AI (GenAI) is widely expected to equalise educational outcomes, yet mounting evidence suggests that its benefits accrue unevenly. This study asks whether the depth of AI use, not merely its frequency, distinguishes high‐performing students and whether AI use reproduces the Matthew effect in academic performance. Using a university‐wide survey of 875 graduating students at a Chinese research university, we constructed a five‐dimension AI use depth index (frequency, weekly time, scenario breadth, verification behaviour and proactive learning) and linked it to objective academic records. Cumulative course grades from the first 3 years preceded capstone scores earned in the final year, allowing a test of whether prior achievement predicts use depth. Higher‐achieving students used AI more deeply on every dimension, a stratification of AI engagement across prior achievement. AI use depth was associated with capstone performance over and above baseline achievement. Capstone performance was 5.05 points higher per standard deviation of AI use depth (95% CI [4.49, 5.61]; β = 0.475), holding prior achievement, discipline, gender and timing of first AI use constant, an increment of 0.211 in explained variance over prior achievement alone. Decomposed models showed that the association was carried by use quality (scenario breadth, verification behaviour and proactive learning), whereas raw use quantity (frequency and weekly time) showed no independent association. The depth‐performance association did not differ detectably across disciplines (interaction Delta R 2 = 0.003; F (6, 859) = 0.822; p = 0.553). Self‐reported gains were strongly associated with baseline achievement but added no information about performance once baseline grades were controlled. The evidence does not show that AI use widens the outcome distribution, and it does not identify the causal direction. The relevant divide is therefore no longer access, or even quantity of use, but the quality and depth of engagement, and AI literacy programmes should consider targeting use quality, verification, breadth and self‐regulated learning rather than usage frequency alone.