AI can narrow the gap between stronger and weaker workers' output, but does it narrow what they can do alone?

This explores whether AI tools narrow the gap between stronger and weaker performers, or only cover it up while the underlying difference in ability stays put. The corpus suggests the answer depends on which gap you mean: the gap in what people produce, or the gap in what they can do on their own. On output, the evidence for levelling is strong. In a randomized experiment, generative AI shrank the advantage that higher-educated adults held on a business problem-solving task by about three-quarters, and lower-education participants kept part of that gain even after the AI was taken away Can AI narrow the education performance gap?. The levelling also shows up in hiring: listing AI skills partly offset the interview penalties faced by older candidates and those without a bachelor's degree, though the effect was much stronger for office assistant roles than for graphic designers Can AI skills help older or less-educated job candidates?.

The picture changes when you look at what people keep once the AI is gone. One study found that workers who used generative AI did much better on content tasks, then showed no improvement when they did similar tasks alone afterward Does AI assistance help workers learn lasting skills?. Another note points to a likely reason: learners without AI ran into more errors and fixed them themselves, and that struggle is where much of the learning happened. AI-assisted learners handed debugging to the AI, and the ones who relied on it most scored lowest on later skill tests Does AI assistance remove a core learning channel through error work?. So AI can close the gap in today's work while removing the experiences that would have closed it for good. That also helps explain why the education study's partial retention stands out, and why it shouldn't be taken as the default.

This matters more because people often can't tell which kind of gain they're getting. Four effects work together to make AI output feel like your own competence: it's unclear who did what, fluent text reads as understanding, thinking gets handed off, and the steps that produced the result are hidden How do AI tools trick users into overestimating their own skills?. Even when a person and an AI working together do better, they capture only about half of the improvement that a better model offers on a given item. Being able to complement each other does not guarantee that they will Why does assisted accuracy capture only half the LLM gain?.

The less comfortable finding is that we may not be able to measure what's happening. Usage data records assisted output well but can't see whether independent skill is building up or wearing away. Current systems watch expertise being used, not expertise being formed, so AI's long-run effect on skill remains genuinely unknown Can we measure whether AI erodes independent skill?. At the level of the whole labor market, the evidence leans toward preserving gaps rather than closing them. Demand for AI skills is clustering in STEM jobs around Python, SQL and machine learning, and non-technical jobs are moving away from that core rather than toward it Is AI creating common skills across jobs or deepening divisions?.

The takeaway: AI is very good at closing performance gaps and much less proven at closing capability gaps, and our usual metrics can't tell the two apart. A workplace could look more equal on every dashboard while the skills that make people independent quietly thin out, most of all for the novices who seemed to gain the most. A side note: several notes in the collection use 'skills' to mean reusable procedures that AI agents store for themselves. They are about agent design, not human skill gaps, so they aren't used here.