Is AI rewarding workers differently than past tech — boosting whoever's results are easy to measure, not just the most skilled?

This explores whether the old story, that new technology rewards the most skilled workers, is giving way to a different one in which what matters is whether a task's output can be measured and checked. The collection has no paper that names or tests 'measurability-biased change' directly. What it does have is evidence that the skill-based story is breaking down, plus a separate set of notes on AI training that helps explain why measurability might be what replaces it.

Start with the skill story. If AI followed the classic pattern, it would widen the gap between strong and weak workers. In a controlled experiment, though, GPT-4 gave the biggest boost to the *weaker* half of consultants and narrowed the gap between them and top performers. There is a catch: when using AI is optional, that effect may reverse Does AI help lower-skilled workers catch up to top performers?. Acemoglu's model points the same way. AI's effect on wages spreads more evenly across demographic groups than earlier waves of automation did, but income still moves from workers toward the owners of capital Will AI increase inequality less than past automation?. So the dividing line is shifting away from skill level toward the split between labor and capital. One result does still look like the old skill story: job-vacancy data from ten countries show demand for AI skills clustering around Python, SQL and data analysis in technical jobs, while other occupations move away from that core Is AI creating common skills across jobs or deepening divisions?.

The measurability idea is easier to see in the training literature than in the labor economics. Models improve reliably only where there is an outside signal that can check the work, such as a judge, a test, a tool's output or a user's correction. Self-improvement without one runs into hard limits Can models reliably improve themselves without external feedback?. On long optimization tasks, the best predictor of agent success was persistence through repeated measure-edit-retry cycles against a benchmark What predicts success in ultra-long-horizon agent tasks?. Even the feedback signals are uneven: a single score tells a system how well it did but loses the information about what to change Can scalar rewards capture all the information in agent feedback?. Put these together and you get a plausible rule: AI becomes capable fastest where outcomes can be scored. That means automation pressure would follow how measurable a task is, not how much skill it takes. This is an inference across notes, not a finding any single paper makes.

Measurability also matters on the hiring side. When AI makes writing cheap, a well-written job application stops being a costly signal of ability. A simulation of Freelancer.com hiring found that removing written signals made hiring 19% less meritocratic, with top workers hired less often and weaker workers hired more Does cheap writing weaken hiring based on worker ability?. So AI does more than automate measurable work. It can also wear away the signals employers use to measure workers in the first place. Firms aren't affected evenly either. Those with more AI exposure replace freelancers faster and more cheaply, which suggests the gains build on a firm's own in-house AI capability Do firms substitute labor for AI at different rates?. Brynjolfsson warns that if AI is designed to automate rather than augment, this can harden into a trap where workers lose bargaining power Does automating labor concentrate power more than augmenting it?.

The honest answer is: not proven, but the pieces line up. The skill-based pattern looks weaker, and the training research gives a reason to think measurability shapes where AI gets good. The open question the collection can't yet answer is the one worth watching: are jobs whose output is hard to score actually more protected, or does AI simply make them easier to score?