Why do firms build automating AI instead of pro-worker AI?

Acemoglu, Autor and Johnson define "pro-worker technologies" as "technologies that make human skills and expertise more valuable," then sort AI into five categories by how they change that value: labor-augmenting (makes workers "more effective at the tasks they already do"), capital-augmenting (makes machines "better, cheaper, or faster" at their current tasks), automating ("substituting machinery or algorithms for tasks that were previously performed by workers"), expertise-leveling (lets "a new set of workers... perform tasks that previously demanded expertise from another domain"), and new-task-creating (which "create[s] new human tasks"). Of the five, only new-task-creating technology is "unambiguously pro-worker," because it alone raises both the quantity and variety of work demanded and the value of the expertise it requires; automating technology is the one category the authors call "unambiguously not" pro-worker. The other three can cut either way, posing "labor market trade-offs, measured in earnings, inequality, and opportunity" rather than a clean verdict.

The authors then ask "why isn't pro-worker AI everywhere?" and answer with an incentive mismatch, not a technical limit: "leading firms perceive greater economic return to building and deploying technologies that automate expertise than those that create new tasks," and the field's "central focus on Artificial General Intelligence (AGI)... also makes it less likely that firms will devote resources to developing pro-worker AI innovations." They frame this as a collective-action gap — "it is not the responsibility of individual firms to make our AI-intensive future work well for workers, but it is in all our collective interest" — and from it derive nine public-policy levers, from sector-targeted grant-making and a DARPA-style prize model to tax-code rebalancing, antitrust, worker-voice mechanisms, "expertise theft" protections, and loosened licensure.

The taxonomy gives a cause to put upstream of the direction-of-effect claim in Does automation raise or lower the skills that remaining work demands?: Autor and Thompson show wages and employment move in opposite directions depending on whether automation strips expert or inexpert tasks from a bundle, but they don't say why firms build automating technology rather than new-task-creating or expertise-leveling technology in the first place. This paper's incentive story — automating expertise pays better than creating it — is one candidate answer. It also sits next to What makes accountable judgment scarce when AI cognition is cheap?, which locates labor-market outcomes in "the social machinery" of verification, liability, and bargaining; Acemoglu, Autor and Johnson name a more specific piece of that machinery — R&D incentives, tax treatment, and antitrust — and argue it is currently tilted toward automation by default, not by technical necessity.

The excerpt offers the taxonomy and the incentive argument but no measurement: it names no firm-level data on how AI investment actually splits across the five categories, no estimate of how much of "current AI focus" really is automation versus the other types, and no test of whether the nine policy levers would shift that split if enacted. The five categories are also offered as analytically distinct but the excerpt does not say how a given deployment (say, an LLM coding assistant) would be classified, or whether real systems cleanly fall into one bucket rather than mixing several. At the strength the excerpt allows, the implication is that the automation-heavy character of current AI investment is a contingent choice shaped by return-on-investment and the AGI framing, not an inevitable property of the technology — which makes it a policy and firm-strategy variable rather than a fixed outcome to forecast around.