Does it matter whether you do the thinking yourself or just supervise the AI — once the AI is gone?

This explores whether the way you engage with a task while using AI, doing the thinking yourself versus supervising what it produces, decides if the benefit is still there when you work alone. The corpus has no single study that tests this directly. Several notes point the same way, though. Gains carry over when AI adds to skills you already have. They carry over much less when AI takes over the effortful part of the work.

Start with where the time goes. AI doesn't so much shrink a task as move the time, away from active work and toward writing prompts and working out what the outputs mean (Does AI really save time, or just change how we spend it?). That shift is the mechanism. The hands-on work you hand off is often the same work that would have built your skill, so time-on-task is a poor way to measure whether you're getting better.

Existing skill matters a great deal. The productivity gains in the studies were measured on tasks inside workers' existing domains. When workers used AI to learn something new, the gains disappeared and the learning suffered (When does AI actually boost worker productivity?). So the boost is real, but it lands on skills you already own and doesn't leave a new one behind. The most direct evidence on independent work is a four-month EEG study of 54 people. Brain connectivity scaled down as reliance on AI went up. Heavy LLM users showed the weakest neural engagement and the poorest memory retention. They also had trouble recalling their own recent work (Does AI assistance weaken our brain's ability to think independently?).

People who lean on AI may not notice any of this. Polished output feels easy to read, and users take that ease as a sign of their own competence (Does processing ease mislead users about their own competence?). The LLM Fallacy is a related error. People credit AI-made output to their own ability, regardless of whether the output was accurate or how much they relied on it (How does AI-assisted work reshape how people see their own abilities?). Someone who was passive can feel capable and so has no reason to practice. The remedy the note proposes is clearer boundaries between human and machine contribution. Better accuracy or forced verification wouldn't fix it.

Engagement isn't simply a matter of more being better. Even correct AI suggestions can hurt reasoning by breaking cognitive immersion, which forces the user to rebuild focus before continuing (Does AI assistance always help reasoning or does it carry hidden costs?). So the real question is where AI enters the work, not whether it does. Reading across these notes, the safest bet is that you keep the reasoning and AI supports it without interrupting. That is an inference, though. The corpus documents the cost of low engagement much better than it shows that high-engagement use preserves transfer.