Human effort in AI-assisted work is assumed to decrease gradually as AI capability increases — better tools, smoother workflow, proportionally less human input needed. The actual scaling has a phase transition.
Note (2026-11-14; read against the paper's full text): the essay says there is a task-specific novelty threshold below which human effort is constant and above which it is linear. The paper's model has no such threshold. Its Proposition 1 says human effort grows linearly with task size for any nonzero novelty fraction, and that it is sublinear only if the novelty fraction and the verification, correction and decomposition costs are all exactly zero; it describes the result as a binary outcome between those two cases, and states that if novelty were a spectrum the sharp transition would soften but the linear scaling would remain. Its simulations show a fitted exponent of about 1.0 for every configuration (novelty fractions 0.1, 0.3 and 0.8), with better agents lowering the coefficient. The paper calls itself a stylized model that does not claim to prove a law of nature, built on five stated assumptions, and its team-size result comes from a simple wall-clock formula in which the optimal team shrinks as the inverse square root of agent throughput; it adds that wall-clock time reaches the square root of task size while total human effort stays linear.
Below a task-specific novelty threshold, AI handles the work and human effort scales as O(1) — constant regardless of task size. Above the threshold, AI cannot handle the novel components and human effort scales as O(E) — linearly with task size. There is no intermediate regime. The transition between constant and linear effort is sharp.
Better AI agents improve the coefficient within each regime but never change the scaling exponent. A more capable AI reduces the constant in O(1) tasks and reduces the slope in O(E) tasks, but the transition between regimes remains discontinuous. The qualitative character of the work — either the human monitors or the human does — is invariant to capability improvement.
The consequence for team design: optimal team sizes decrease as agent capability increases. More powerful AI means fewer humans, not the same number of humans working faster. The human role shifts from distributed execution to concentrated evaluation at the novelty boundary. The bottleneck is not effort quantity but effort type — novel judgment that cannot be parallelized across more people.
The structural observation: AI capability improvement does not smoothly reduce human effort but instead moves the location of a cliff. Everything below the cliff becomes trivially automated; everything above it remains fully human. The cliff moves, but its shape does not soften.