Can metacognitive feedback stop students from offloading to AI?

In a preregistered online experiment (N = 704) with a 2×2 design and a no-AI control, participants practiced fraction arithmetic with an LLM-based assistant that "provided solutions only on explicit request," then took an unaided test. Metacognitive feedback that "makes the implications of offloading for users explicit" reduced answer offloading (OR = 0.47) and improved test performance (OR = 1.51). The second intervention, an effort-based reward that "incentivizes less extensive LLM assistance," showed no evidence of affecting either outcome.

The paper starts from the worry that offloading "can reduce opportunities to practice skills," and asks how to prevent deskilling "without restricting access to AI." Its answer aims at the moment of handover rather than at availability. The feedback appears to help learners "regulate their use of assistance while leaving the full range of LLM support available." Exploratory analyses point to where the risk sits: participants who requested complete answers more often tended to do worse on the test, and answer offloading "varied strongly across participants," more common among those lower in need for cognition and perceived confidence. The authors conclude that "the relevant risk lies less in access to LLM assistance itself than in how much cognitive work learners choose to hand over." They also report no evidence that access to the assistant impaired unaided performance under this design.

The framing matches Does AI assistance actually harm the way developers learn?, where the variable that matters is how the assistant is used, not whether. That study sorted observed behaviors into low- and high-scoring patterns. This paper adds an intervention on the behavior itself, and a measured effect from moving it. The finding also sits against the introduction's own examples of skills eroding after AI-assisted practice, and against Does AI assistance help workers learn lasting skills?. Here the AI arm shows no evidence of an unaided deficit, so the contrast may lie in how much answer-taking the design invites, though the excerpt does not test that. The direction of the exploratory result, heavy answer-requesters scoring lower, is consistent with Does AI assistance weaken our brain's ability to think independently?, though that study measures neural connectivity and this one measures test performance.

The excerpt is silent on several things that limit how far this travels. It does not say what the feedback said, when it appeared, or whether it was repeated. It does not say whether the effect persists past a single online session or extends beyond fraction arithmetic. It does not explain why the reward failed, and "no evidence" is not evidence of no effect. The exploratory links between offloading, test scores, need for cognition and confidence are described only as tendencies, so the causal direction is open. It also reports nothing on what the feedback costs the learner, which matters given Does AI assistance always help reasoning or does it carry hidden costs?. What follows at this strength: for anyone designing learning assistants, the handover decision is the lever to test first, and the null AI-versus-control result should be read as belonging to an assistant that gave answers only when asked.