AI as Normal Technology (2025)

Arvind Narayanan and Sayash Kapoor argue that artificial intelligence should be framed as "normal technology" - a powerful, general-purpose tool continuous with past technologies like electricity and the internet, not a separate, potentially superintelligent species. They present this stance as description, prediction, and prescription: AI will remain human-controlled, its transformational effects will unfold over decades, and policy should assume gradual invention → innovation → adoption → diffusion across distinct timescales. Evidence emphasizes slow diffusion into safety-critical domains: many consequential predictive systems still use simple, interpretable regressions rather than modern transformers; high-profile failures such as Epic’s sepsis model, early Bing “Sydney” conversations, and Google’s Gemini image errors illustrate testing and deployment gaps. Safety constraints and regulation (FDA, EU AI Act) impose practical speed limits on risky applications.

From this baseline, the authors map a likely division of labor in which humans and organizations exercise primary control and an increasing share of work becomes “AI control.” Risk analysis shifts accordingly: accidents, misuse, arms races, and misalignment call for reducing uncertainty and building resilience rather than extraordinary measures against imagined superintelligence. They warn that drastic interventions premised on fast takeoff will worsen outcomes if AI behaves like other capitalist-era technologies, whose harms - especially inequality - require institutionally grounded responses. The essay positions these claims as a median forecast to guide policy and further research.