Does AI assistance help workers learn lasting skills?

The standard implicit theory of AI-augmented work is that working with AI is a form of practice — the worker performs tasks, learns from the AI's contributions, and gradually internalizes capabilities they can later deploy independently. The implicit theory predicts that AI-assisted performance should produce skill transfer: workers who used AI on Task A should perform better on Task B (similar but unassisted) than workers who did not have AI on Task A.

Wu et al. measured this directly. Workers used generative AI to perform content-creation tasks (drafting Facebook posts, performance reviews, welcoming emails) and showed substantial immediate-task improvement compared to controls. The transfer prediction failed. When the same workers performed similar tasks independently afterward, their independent performance was not improved. The skill that would have justified the implicit-theory expectation did not transfer.

This is empirically distinct from the "AI as exoskeleton" framing in an important way: exoskeleton describes performance disappearing when AI is removed in the same task. Wu et al.'s finding describes performance disappearing across tasks, even when the second task is similar to the first. The non-transfer is between contexts, not just within a single use. The AI-assisted experience does not function as practice for unassisted work in the way that unassisted experience would.

The diagnostic implication for skill development is significant. AI-assisted work is not a substitute for unassisted practice when the goal is durable skill. Workers who only ever produce content with AI assistance accumulate a portfolio of AI-assisted outputs and no underlying skill they can deploy without AI. The mismatch between accumulated portfolio and accumulated skill is invisible to the worker and to organizations that measure portfolio rather than capability.

This bears on professional development design under AI. Practice models calibrated to pre-AI work assumed that doing the work was learning the work. Under AI, this no longer holds. Practice for skill-formation may need to be deliberately structured to exclude or constrain AI assistance — making explicit what was implicit when work and learning automatically coincided. Apprenticeships, training programs, and education will need to re-introduce structured unassisted practice if skill development is to continue.

Does AI assistance build lasting skills or temporary abilities? is the within-task companion; this is the across-task version. Together they describe a system where AI-assisted performance is decoupled from skill in both temporal directions: it does not require skill to produce, and it does not produce skill as a byproduct.

The strongest counterargument: longer time horizons and explicit reflection might produce transfer. Possible, but the experimental design controlled for immediate practice; if transfer requires explicit additional steps beyond AI-assisted work, the implicit theory of "AI use as practice" is already wrong. Adding deliberate skill-formation steps reintroduces the unassisted practice the implicit theory was meant to replace.

Enrichment — the skill-revaluation channel into inequality. An interdisciplinary inequality review reframes this non-transfer finding as one of the mechanisms by which generative AI can deepen disparity rather than reduce it. The review observes that generative AI does not merely alter practices but "fundamentally transforms the valuation of knowledge and skills" — as AI begins to exceed human skill at, say, written communication, the incentive to master syntax, vocabulary, and grammar wanes, devaluing foundational learning. The within-task gain that does not transfer is exactly how this incentive erosion operates: the worker gets the output without acquiring the skill, so the rational move under time pressure is to keep delegating and never learn. At a population scale this widens gaps rather than closing them, because those who already hold the foundational skill can supervise and correct AI output while novices who skipped acquisition cannot — making the non-transfer effect a distributional hinge, not just an individual-development concern. Source: Social Theory Society — "The impact of generative artificial intelligence on socioeconomic inequalities and policy making", https://doi.org/10.1093/pnasnexus/pgae191