Showing people how much an AI is doing their thinking helped in one experiment, but does it last?

This asks whether metacognitive feedback (telling people how much they're offloading to an AI, and what it costs them) has effects that outlast a single online session. The corpus can't answer that yet. The one direct study is a 704-person preregistered experiment: feedback that highlighted offloading costs cut answer requests to an LLM in half and raised unaided test scores by 51%, while an effort-based reward did nothing measurable Can metacognitive feedback stop students from offloading to AI?. Those results are measured within the session. The note reports no delayed test, no return visit, and no check on what happens once the feedback is removed.

The corpus does show why the time scale matters. A four-month EEG study of 54 people found that brain connectivity scaled down as AI reliance grew. LLM users showed the weakest neural engagement and the poorest memory, and they struggled to recall their own recent work Does AI assistance weaken our brain's ability to think independently?. So offloading builds up over months, while the feedback has only been tested over one sitting. That study didn't test any intervention, so it can't say whether feedback would slow the decline. It does suggest that a lasting fix would have to work at the same slow pace as the problem.

Some lateral evidence suggests why the feedback worked, which bears on whether it might last. Explaining the cost beat rewarding effort, and something similar happens in AI training: natural language critiques that explain why an attempt failed broke performance plateaus where numerical rewards couldn't, because a score doesn't say what went wrong Can natural language feedback overcome numerical reward plateaus?. In an 80-person lab study, an assistant that asked reflection questions alongside its advice beat one that only advised Do reflection questions help people make better decisions with AI?. If explanation is what changes people, it might stick more than a reward would. But this is an analogy, and nobody in this library has tracked whether a person's habits actually shift.

Two cautions point the other way. Feedback about AI use is itself an intervention, and even correct AI interventions can break cognitive flow and force people to rebuild focus Does AI assistance always help reasoning or does it carry hidden costs?. Whether repeated feedback would carry that cost over weeks is untested. And for AI agents, metacognitive loops that humans design and impose from outside fail when conditions shift, and improvement lasts only when the agent builds its own Can AI systems improve their own learning strategies?. The human parallel is open: feedback delivered from outside might work only while it is present, unless it teaches people to monitor themselves. A follow-up that tests unaided performance weeks later, with the feedback removed, would settle the question, and the library doesn't hold one yet.