Once the AI is taken away, do people keep the gains they got with it, or was the boost only on loan?

This explores whether quality gains from AI assistance stay with people after the tool is gone, or whether they disappear along with it. The honest answer from this collection is that no study here tests that directly. None of these notes takes the tool away from people and measures what's left. The collection is still useful, though, because it explains why the question is so hard to answer and why people's own sense of the answer may be wrong.

Start with the measurement problem. Usage data captures what people produce with AI, not what they can do without it. One note calls this a gap between expertise in use and expertise forming: current systems are good at seeing the first and nearly blind to the second, so AI's effect on skill remains undetermined Can we measure whether AI erodes independent skill?. The well-known productivity results only measure people while they're assisted. Support agents resolved 15% more issues per hour with AI, and the least experienced gained the most in both speed and quality Does AI assistance help less experienced workers most?. Whether those newer agents absorbed the better habits or were simply carried by the tool is exactly what that kind of data can't show.

The second problem is that people misjudge their own skill. The 'LLM fallacy' describes users counting AI-assisted work as proof of their own ability, especially when the output is so smooth that you can't see where the human stopped and the AI started Do AI-assisted outputs fool users about their own skills?. Self-reports are already unreliable on this. Experienced developers expected AI to speed them up by 24% and were actually slowed by 19% Do AI coding tools actually speed up experienced developers?. If people can't judge their speed while using the tool, their sense of what they'd keep without it deserves even less trust. Workplace interviews point the same way: effort and uncertainty disappear into polished deliverables, so the finished product gets treated as proof that the learning happened Which workplace cues survive AI mediation and which disappear?.

One result comes close to a removal test, and it's a surprising one. At ICML 2026, banning LLM use in peer review rather than allowing limited use barely changed scores, decisions or reviewer confidence Does banning LLM use in peer review change review outcomes?. One reading is that whatever reviewers gain from AI doesn't depend on having the tool in hand. But many reviewers broke whichever rule they were given, so it's a blurry picture of 'without AI'. It also reveals something else: in practice, taking the tool away is harder than it sounds.

Here's the takeaway you might not have expected. The groups AI helps most, newcomers lifted toward expert-level output, are also the groups for whom 'did they learn, or were they carried?' matters most and is hardest to see. Meanwhile, the most experienced support agents saw slight quality declines Does AI assistance help less experienced workers most?. That's a reminder that gains and losses can sit side by side even while the tool is in use. To answer this question properly, someone needs to run the experiment the collection is missing: assist people, then take the tool away and measure again.