What makes accountable judgment scarce when AI cognition is cheap?

The paper rejects "Will AI destroy jobs?" as "a low-resolution question" because it treats a job as an indivisible object and AI as a uniform substitute. Its replacement is a framework that treats generative AI as "cheap, scalable, and fallible cognition" and asks how that changes the task composition of production, the formation of expertise, and the allocation of rents. The conclusion names the scarce asset: "accountable judgment." Human work "survives and flourishes" where people ask consequential questions, recognize context, evaluate machine output, persuade other humans, absorb legal and moral accountability, and learn from consequential practice.

The mechanism is a split between technical reach and realized displacement. The abstract separates what large language models can do from equilibrium labor-market displacement, using a task vulnerability index and an adoption condition that makes "verification, liability, trust, and governance explicit." Occupations are then modeled as "governance bundles rather than task lists," meaning tasks bound together with responsibility, sequencing, memory, tacit knowledge, customer attachment, and learning. The analyst or paralegal is "a workflow node," not a list of inputs. Net labor-market effects come out as a balance among task compression, scale expansion, new human work, and institutional bargaining. The conclusion goes furthest: "the social machinery built around" models "matters more" than raw capability, and the machinery it lists is workflow design, question rights, apprenticeship, verification, liability, education, competition, bargaining, public measurement, and rent allocation.

Against the nearest notes, this is the institutional counterpart to a mechanism story. Does AI reshape expert work into knowledge management? treats evaluating AI output as what is left when the interested work of thinking is hollowed out. This paper also makes verification central, but pairs it with accountability and question selection and says it can compound, on one condition: institutions must preserve "the learning, question-selection, authority, and bargaining conditions." The two agree on where the residual role sits and differ on whether it is a trap or a foundation. Does incremental AI replacement erode human influence over society? describes the same fork at societal scale, where "abundance with hierarchy" is the disempowerment outcome. Does concentrated AI exposure enable workers to adapt and reallocate? is an empirical instance of the task-level view this paper argues for, since it measures exposure by task rather than by job. Does AI assistance help workers learn lasting skills? bears on the apprenticeship margin, because gains that do not persist are what a learning-preserving institution would have to guard against.

The excerpt is only the abstract, one introduction passage, and one conclusion, so it establishes a framework and a position, not findings. It does not state how the task vulnerability index is built, what data or calibration the adoption condition uses, or whether any occupation or firm was measured. The claim that institutions matter "more" than capability is asserted in the conclusion and is not shown in the excerpt. What follows at that strength is a set of margins to check, namely verification, liability, apprenticeship, and rent allocation, and not a forecast for any occupation.