Does ChatGPT help writers think less deeply about revision?

In a randomized experiment, 117 university students completed a two-stage English reading-and-writing task under one of four conditions: no support (CN, 30), ChatGPT 4.0 restricted to the task topic (AI, 35), a human writing expert (HE, 25), or a writing-analytics checklist tool (CL, 27). The authors report that "ChatGPT group outperformed in the essay score improvement but their knowledge gain and transfer were not significantly different" from the other groups, and that "learners who received different learning support showed no difference in post-task intrinsic motivation." Process-mining of the revision stage found the AI group "exhibited SRL processes closely tied to the interactions with ChatGPT" and showed "relatively fewer metacognitive processes (e.g., evaluation and orientation)" than the HE and CL groups, while human experts "triggered more metacognitive process associations (e.g., transitions between orientation and evaluation)."

The paper names this pattern metacognitive laziness, defined as "learners' dependence on AI assistance, offloading metacognitive load, and less effectively associating responsible metacognitive processes with learning tasks." It grounds the construct in Risko and Gilbert's cognitive offloading, where reliance on an external resource reduces engagement in monitoring and self-regulation, and in Alter et al.'s account of cognitive disfluency, where friction provokes deeper analytical processing and its absence removes that prompt. The authors read the essay-score gain alongside the absent knowledge-transfer gain as the two faces of the same mechanism: ChatGPT improved the artifact learners handed in without improving what they could do without it.

This sharpens Does ChatGPT harm informal learning compared to Google Search?, which traced reduced agency and higher meta-cognitive load to a chat interface absorbing search's selection and evaluation steps; here the comparison condition is not a results page but two other active forms of support, so the gap is not chat-versus-list but ChatGPT-versus-expert and ChatGPT-versus-checklist, and it shows up specifically as missing evaluation and orientation processes during revision rather than as a test-score deficit. It parallels Does ChatGPT help students code better but remember less? in splitting a performance gain from a learning-depth cost, and it extends Does AI assistance actually harm the way developers learn? by locating the low-engagement pattern inside a controlled four-arm comparison rather than inferring it from clustering developer behaviors. It complements Can metacognitive feedback stop students from offloading to AI?: that paper intervenes on offloading with a feedback prompt and measures a recovered test score, while this one documents the uncorrected baseline — an AI condition with no such prompt, where offloading shows up as fewer SRL transitions rather than a tested remedy.

The study does not establish that ChatGPT caused less metacognition in some general sense: it ran one writing task, in a lab, over roughly three hours, with group sizes of 25–35, and the authors themselves attribute some null results to limited task duration and sample size. It also cannot separate the chat interface from the specific restriction that ChatGPT "was limited to providing advice rather than directly generating an essay," so the result speaks to that configuration, not to generative AI assistance in general. At the strength this evidence allows, the finding is a within-study contrast: among four comparably scaffolded supports, only the AI condition paired a short-term performance gain with reduced evaluative and orientation activity during revision, which is the condition under which the authors judge metacognitive laziness most likely to set in.