Lab vs. life: Dissecting “AI as normal technology”

Abstract

A companion to the author’s dissection of AI 2027, this piece examines Arvind Narayanan and Sayash Kapoor’s AI as Normal Technology and teases out the assumptions that yield its forecast of transformative impact on the timescale of decades. On that paper’s account, progress moves through invention, innovation, adoption and diffusion, with each stage depending on feedback that can only be gathered through real-world use — the difference between Waymo’s two decades of field trials and AlphaZero’s three hours of self-play. Newman draws out the load-bearing claims: that benchmark performance systematically overstates real-world capability, since the easier a task is to measure the less it resembles the complex, contextual work of professional practice; that deployment is slowed by trade secrets, tacit organizational knowledge and privacy constraints; that adoption meets safety and regulatory friction, with generative AI’s roughly 40% adoption rate covering only 0.5–3.5% of actual work hours; that irreducible stochasticity caps AI performance in domains such as geopolitical forecasting and persuasion; that humans will remain in a supervisory role for task specification and safety; and that greater intelligence does not translate automatically into greater power. He closes with the counterarguments fast-timeline proponents would raise. – AI-generated abstract.