AI levels tasks that have a template, not judgment; the leveled work gets cheaper — resolves 2027-07-17

Several studies find AI helps weaker performers most and narrows the gap, and the question is whether that makes workers more equal or lets a few stars take everything.

Companies use the answer to decide whether to hire less experienced people and add AI, cut staff, or train a few power users, which moves pay and training budgets.

Three views compete, AI as leveler, as an escalator that lifts everyone equally, and as a kingmaker for a few, and SecondSource reads the answer as different at each layer.

SecondSource’s reading: skill leveling has to be read in three layers. At the task level it holds: in work with a quality template to check against (writing, ideas, legal, customer service, consulting analysis), AI pulls the bottom half toward the average, backed by five peer-reviewed studies.

At the judgment level it reverses: in work with no template, AI widens the gap. A field randomized trial in Kenya (Otis and co-authors, Management Science, 2025) found strong performers improved 15% while weak performers fell 8%.

At the market level, the leveled segment of work gets commoditized and cheaper. That layer is our inference and has not been directly tested, so we set ourselves a 12-month observation window.

One Useful Thing (Ethan Mollick, Wharton)
“AI as Leveler — We already know one major effect…”

Reviewed July 2027: (1) a clean leveling RCT reproduced in a domain without a template revises the judgment leg; (2) an entry-level decline of equal size in augmentation-type occupations refutes the by-domain commoditization reading; (3) an AI wage premium still widening after controlling for role and tenure upgrades the narrow king-making claim. If none of the three occurs, the directional call is a hit and is registered as a formal thesis.

What would overturn it

Any of the three triggering rewrites the corresponding leg. The next Canaries release and Yale's quarterly reports are the data source for the second. Direct time-series rate data would allow the price-convergence inference to be closed out early.