Where do human experts focus as AI capabilities expand?

Agrawal, Gans and Goldfarb's NBER working paper (not peer-reviewed) models an economy of routine workers, human geniuses and AI geniuses answering questions that vary by distance from a known point. Routine workers apply "existing knowledge to form conjectures," while geniuses "generate new but ephemeral knowledge... working from first principles" at a cost that rises with distance from what is known. A manager routes each question by comparative advantage, a rule the paper treats as "a key learned task – from experience." Before AI, scarce genius capacity goes first to the edges of the routine domain, "where the incremental value over routine is greatest." Adding AI geniuses changes this: "In the short run, human geniuses specialise in questions that are furthest from existing knowledge, where their comparative advantage over AI is greatest. In the long run, routine workers may be completely displaced if AI efficiency approaches human genius efficiency."

The mechanism follows Amodei's (2024) "country of geniuses in a data center": AI geniuses "approach problems in a similar manner" to humans but may be less efficient (a cost parameter ηAI ≥ η), and are available in "effectively unlimited supply" while human geniuses stay scarce (LG < 1). Because the manager's routing rule is learned rather than instantly re-optimized, the short run is inertia: AI geniuses get allocated alongside human geniuses to genius-type questions, pushing humans toward the most novel ones, where their edge over AI is largest. Only once the manager "adjusts their routing rule" does AI move into the routine domain and, if its efficiency closes in on human genius efficiency, "displace routine work." The conclusion flags the model's limit: it treats AI geniuses as "independent agents that replace human workers," and calls instead for modeling AI as a complement to human "contextual understanding, ethical judgment, and creative insights."

This gives a model-based timeline to the empirical question Does automation raise or lower the skills that remaining work demands? poses: here it is distance from the knowledge frontier, not task removal from a bundle, that decides whether a worker's value rises or falls. It also relocates the scarce asset named in What makes accountable judgment scarce when AI cognition is cheap?: this model puts scarcity at the frontier of existing knowledge rather than in accountability, though both hinge on the same routing decision — "knowing who to ask." Against Why do firms build automating AI instead of pro-worker AI?, AI geniuses fit none of the five categories cleanly, since they create knowledge rather than automate or augment fixed tasks.

The excerpt draws from the introduction, method, discussion and conclusion of a working paper that has not been peer-reviewed, with no calibration, data or empirical test — ηAI, LG and the routing thresholds are theoretical parameters. The paper itself cautions that modeling AI geniuses as "independent agents that replace human workers" is a choice, not a claim about actual deployment, and names "artisans, artists or persuasive geniuses" as categories it has not modeled. What follows at this strength is a framing question for any AI-adoption claim: whether the AI in question acts as an independent frontier-level agent or as a complementary tool — the model only speaks to the former.