How does generative AI actually change worker skills?

Researchers at TU Delft conducted "semi-structured interviews... with 38 professionals in the Netherlands, including academics (e.g., lecturers and professors) and non-academic professionals (e.g., consultants, analysts, engineers, legal professionals, and public sector employees) with varying levels of experience." Inductive thematic analysis of how these workers described skill-related change surfaced four outcomes rather than two: "skill development, skill maintenance, skill erosion, and skill revaluation." The paper states its target directly: "Current debates typically frame GenAI's impact on skills in binary terms—upskilling versus deskilling—yet empirical evidence on how workers themselves experience these changes in everyday practice remains limited."

Each outcome names a distinct mechanism in the workers' own accounts. Skill development came from "learning from GenAI outputs, expanded information access, and offloading routine tasks to focus on higher-level work." Skill maintenance described workers who perceived "little or no change," which the paper links to "selective and critical use" rather than to avoidance. Skill erosion is "diminished ability to perform tasks independently without GenAI support." Skill revaluation is a shift in "perceived skill importance as certain tasks became delegable while others gained prominence." The paper's summary claim is that impact is "heterogeneous and practice-dependent" — the same technology produces different outcomes depending on how a worker uses it, not a uniform direction of change.

This four-outcome framework gives a taxonomy for findings that the library otherwise holds as separate, mechanism-specific claims. Does generative AI prevent juniors from getting entry-level work? and Does AI turn freelance work into validation instead of creation? both describe cases that this framework would classify as skill erosion, tied to a changed task mix rather than individual choice. Does AI assistance help workers learn lasting skills? is consistent with erosion risk at the task level. None of those papers name maintenance or revaluation as separate categories; this excerpt's contribution is the broader typology and its explicit rejection of the binary framing those other papers' erosion/deskilling language can imply.

The excerpt is an abstract: it gives no breakdown of how many of the 38 participants fell into each outcome, no detail on which professions clustered where, and no account of what "selective and critical use" consists of in practice. It is a self-reported perception study, not a measured-skill study, so the four outcomes describe how workers interpret change, not verified changes in capability. The implication the abstract supports is narrower than a general claim about AI and skill: perception of skill effect depends on how a worker uses the tool and what outcome category they land in, so interventions or further research aimed at "skill effects of AI" should specify which of the four outcomes they mean.