They take BLS/ACS education attainment by detailed occupation, then estimate (from Claude usage) the average years of schooling needed to understand tasks/prompts associated with each occupation - and plot that against % of workers with a BA+. The relationship is directionally clear - the occupations whose prompts look “more educated” tend to be the ones with more degree-holders.

This puts downward pressure on the Equality dial. If the ability to “use” AI is gated by “prompt literacy” (which itself tracks educational attainment), then AI becomes a mirror that reinforces existing advantage - higher-credential workforces extract more value, faster, while everyone else gets a thinner slice of the productivity upside.

This is the uncomfortable version of “AI for everyone” - it may be widely available, but not equally legible. And that’s how unequalising effects compound.

> The interface layer is thickening. If you disagree with my interpretation, or you’ve spotted a better signal then reply and tell me.