Bounded AI help beat open AI help when the AI was taken away

Xinran Chen of Jinggangshan University ran 180 undergraduates through eight weeks of academic writing under three conditions: no AI, bounded AI support with a mandatory reflection step, and open AI collaboration. 168 finished. The measurement that matters came at week eight, on a supervised task with no AI available.

Open collaboration won while the tool was there and lost when it was not. On the independent task the bounded-support group scored higher on writing quality (+0.27), higher-order thinking (+0.35), argument depth (+0.42) and independent revision quality (+0.39). A mediation model put the mechanism in the middle: deeper offloading predicted weaker independent performance through reduced higher-order thinking, indirect estimate −0.34, 95% CI [−0.48, −0.21].

The honest caveat is in the paper’s own numbers. Wild-cluster p-values for those contrasts run from .050 to .063 — on the line, not past it, with six intact classes rather than randomised individuals. This is a directional result from one cohort in one course, not a settled effect.

What makes it worth recording is the design. It does not ask whether AI helps; it asks what is left in the person afterwards, and it separates the two by taking the tool away.