Skill Distribution and Who Benefits
The consistent finding that gains concentrate among lower-skilled workers, the competing explanations for it, and the conditions under which the pattern reverses.
The most consistent result across generative AI productivity studies is distributional rather than aggregate: lower-performing and less experienced workers gain most, and top performers gain little or nothing. It appears in customer support, in professional writing, and in consulting-task experiments, and it is a reversal of the pattern for most previous information technology, which was skill-biased and widened the gap.
The competing explanations
Access to codified best practice. The model encodes what a good response looks like, so a novice receives guidance that previously required experience or a mentor. On this account the model is a delivery mechanism for tacit knowledge, and it explains why the effect is largest in tasks with a clear notion of a good answer.
Ceiling effects. A top performer is already near the achievable maximum on the measured dimension, so there is little room for improvement in the metric even if the model helps them qualitatively. This account predicts the effect would shrink if the measurement captured quality above the current ceiling.
Task composition. Experienced workers spend proportionally more time on the parts of the job the model does not help with, coordination, judgement, exceptions, so a smaller share of their time is exposed. This account predicts the pattern reverses when the model becomes capable at those parts.
The three make different predictions and are not currently distinguished by the evidence, which is worth stating rather than picking one.
Where the pattern reverses
The METR programming result is the clearest counterexample: experienced developers working in repositories they knew well were slowed. The mechanism proposed is that their context was the scarce input, and reviewing a suggestion made without that context cost more than writing the code.
This suggests a boundary condition rather than a contradiction. The levelling effect holds where the model's knowledge exceeds the worker's, and it inverts where the worker's context exceeds the model's. As models gain context, through longer windows, better retrieval and repository-level understanding, the boundary moves, and which side a given worker sits on is not fixed.
When it breaks
Levelling in measured output is not levelling in outcomes. If a firm responds to uniform performance by paying less for experience, the compression in measured productivity becomes compression in wages, which is a different and less benign outcome than the productivity result alone suggests.
Learning may be affected. If novices produce expert-level output without acquiring expert judgement, the pipeline that produces future experts is disrupted. There is not yet good longitudinal evidence on this, and it is the concern most often raised and least often measured.
The measured dimension may be the wrong one. Studies measure what is measurable: speed, resolution rate, graded quality on a rubric. If experienced workers' advantage lies in the cases that fall outside the rubric, the finding that they gain nothing is a statement about the rubric.
Effects at the individual level need not survive at the firm level. A firm whose novices now perform like experienced staff may hire differently, restructure teams, or change what it sells. The individual-level result is an input to that and does not determine it.
12 flashcards for this concept
Click a card to reveal the answer.