The Novelty Bottleneck: A Framework for Understanding Human Effort Scaling in AI-Assisted Work

Human effort doesn't scale smoothly with AI improvements: The paper finds that human effort required in AI-assisted work transitions sharply between two states rather than gradually improving. This means that as AI systems get better, organizations shouldn't expect a smooth, predictable reduction in human workload—instead, improvements may suddenly make certain tasks feasible or remain stubbornly difficult.

Better AI agents don't fundamentally change the workload structure: While more capable AI systems can reduce the amount of human effort needed (the "coefficient"), they cannot eliminate the core bottleneck created by novel or unprecedented decisions. This suggests that even advanced AI won't solve the fundamental problem of human judgment being required for new challenges.

The "novelty fraction" is the critical factor: The paper identifies that the proportion of a task requiring genuinely new human judgment—rather than applying existing knowledge—is the key parameter determining how much human oversight is needed. Organizations should focus on measuring and understanding this novelty fraction rather than just overall task complexity.

AI creates an asymmetric safety profile: The framework reveals that AI systems face bottlenecks when applied to frontier research and novel problems (where human judgment is essential), but are "unbottlenecked" when exploiting existing knowledge. This means AI poses different risks in different contexts—less risk when constrained by human oversight on novel work, but potentially greater risks when deployed at scale on routine tasks.

Optimal team sizes shrink as AI improves: Counterintuitively, organizations may need fewer human workers as their AI agents become more capable, though the total human effort required for novel work remains substantial. This has significant implications for workforce planning and the distribution of AI benefits.