When does AI actually boost worker productivity?

The reigning empirical story about AI in the workplace is that AI produces large productivity gains, especially for less-experienced workers (Brynjolfsson 15%, Dell'Acqua 12.2%, Peng 55.5% on coding). The natural extrapolation is that AI is most valuable where existing skill is lowest — which would make it especially valuable for novices and learners.

The Skill Formation study breaks this pattern. When developers used AI to learn a new asynchronous programming library — rather than apply existing programming skills — the productivity gain disappeared. Average completion time was not significantly different from the control group. The aggregate gain hid heterogeneity: a small subset (about 20%) who fully delegated coding to AI completed faster, but the majority who tried to use AI as a learning aid spent more time interacting with the AI than they saved on the coding.

This matters for how prior productivity findings should be interpreted. The famous gains were measured on tasks where workers already had the skill; AI sped up the application of skill. The Skill Formation study measured a different task — acquiring the skill in the first place — and the gain vanished. Different studies were measuring different things, and the productivity story does not generalize across them.

The diagnostic implication is significant for organizational AI deployment. Tasks that involve applying existing skill at speed will see real productivity gains; tasks that involve workers learning unfamiliar territory will not, and may impose new costs in time and skill formation. Organizations that deploy AI uniformly across both task types are misallocating — they will get gains in the first category and losses in the second, with the aggregate appearing more positive than the disaggregated picture would.

It also bears on how junior workers should be deployed. The "AI helps novices most" story applies to novices doing familiar work; for novices doing unfamiliar work, AI may produce neither productivity nor learning. The right deployment of AI to junior workers requires distinguishing between these two task types in real time — a managerial competence that does not yet have practice patterns built around it.

The strongest counterargument: agentic tools and better interfaces will eventually deliver gains even on learning tasks. Possible at the limit, but the mechanism would be different — AI doing the work entirely, with the worker not learning at all — which closes the productivity gap by closing the learning channel rather than improving it.