AI as normal technology is a framework proposed by Arvind Narayanan and Sayash Kapoor (Knight Columbia, 2025) holding that AI should be understood as a transformative but normal general-purpose technology — not a separate species — subject to the same dynamics of slow diffusion, organizational adaptation, and institutional change as prior general-purpose technologies (AI as Normal Technology). It rejects superintelligence framing, predicts decades-long diffusion, and advocates resilience over drastic intervention.
Core claim
Narayanan and Kapoor present "AI is normal technology" as three claims at once: a description of current AI, a prediction about the foreseeable future, and a prescription about how AI should be treated. The framework rejects technological determinism — especially the notion of AI itself as an agent in determining its future — and emphasizes continuity with past technological revolutions (AI as Normal Technology).
Key concepts
Methods–applications–diffusion distinction. Rapid progress in AI methods does not translate to rapid applications or diffusion; each stage has independent speed limits. Benchmarks measure methods but are routinely misinterpreted as measuring applications, a construct-validity problem the authors identify as a driver of AI hype.
Capability versus power. Narayanan and Kapoor argue that "intelligence" is not the right concept for analyzing AI impact; what matters is power, the ability to modify one's environment. On this account humans are already "superintelligent" compared to pre-technological humans through tools rather than biology, and AI is the latest such tool.
Many flavors of control. The control problem is not limited to "alignment" or "human-in-the-loop." The authors point to existing approaches including auditing, monitoring, fail-safes, circuit breakers, least-privilege access, formal verification, and redundancy, and argue that an increasing percentage of human jobs will consist of AI control.
Resilience over drastic intervention. The framework holds that drastic policy interventions premised on the difficulty of controlling superintelligent AI will make things worse if AI turns out to be normal technology, and that reducing uncertainty should be a first-rate policy goal.
Debates and positions
The framework stands in direct contrast to acceleration- and superintelligence-framed views associated with Dario Amodei and with Sullivan-Feldman. The two camps differ across several dimensions:
| Dimension | AI as Normal Technology | Amodei / Sullivan-Feldman |
|---|---|---|
| Timeline | Decades of diffusion | 1–2 years to powerful AI |
| Superintelligence | Incoherent as usually conceptualized | Plausible enough to plan for (Amodei); one axis of uncertainty (S&F) |
| Key risk | Inequality, concentration of power | Existential: autonomy, bioweapons, authoritarianism |
| Policy | Resilience, sectoral regulation, reduce uncertainty | Transparency-first, then targeted intervention; export controls |
| Control | Many existing flavors; tractable | Requires massive new science |
| Benchmarks | Don't measure real-world utility | Scaling laws as predictive of capability |
Which view holds depends on empirical questions that are not yet resolved. If Narayanan and Kapoor are correct, much of the Eight Worlds Framework's matrix becomes irrelevant because the superintelligence axis collapses; if Amodei is correct, the normal-technology framing is, on his account, dangerously complacent. Simon Willison's practitioner observations in his year-in-review posts (year-in-review posts) describe rapid progress in methods alongside a real capability-deployment gap.
Survey evidence has been cited as bearing on the slow-diffusion prediction. A Gallup survey conducted for the Walton Family Foundation and GSV Ventures, covering more than 1,500 U.S. respondents ages 14–29 (February–March 2026), found Gen Z adoption plateauing despite increasing access, with roughly 50% reporting daily or weekly use, flat year-over-year. Attitudes also declined: hopefulness fell from 27% to 18%, nearly a third reported that AI makes them feel angry, and close to half of working Gen Z respondents said the risks outweigh the workplace benefits, an 11-point increase (Source: nytimes.com).
Knight Columbia 2026 follow-on essays
Three Knight Columbia 2026 essays explicitly engage Narayanan and Kapoor and together form an alternative to AGI- and Singularity-framed AI policy analysis:
- Anticipatory AI Ethics (Lazar, Knight Columbia, May 1 2026) (Seth Lazar, Anticipatory AI Ethics, May 1 2026) engages on condition 2, the research-to-deployment gap. Lazar argues the gap is narrower than Narayanan and Kapoor claim, citing OpenAI's 400M+ monthly active users, Microsoft Copilot's OpenAI integration, Alexa+ running on Anthropic Claude, and a "vessels waiting to carry capable models to market" pattern. Lazar frames this as a contested empirical claim rather than a hard contradiction.
- AI as Social Technology (Farrell + Shalizi, Knight Columbia, May 11 2026) (Henry Farrell and Cosma Rohilla Shalizi, AI as Social Technology, May 11 2026) is a complementary framework. Both reject AGI/Singularity framing; Farrell and Shalizi add the institutional-organizational dimension they say Narayanan and Kapoor under-emphasize, treating LLMs as coarse-grainings that reorganize social relationships, continuous with markets, bureaucracies, the price mechanism, and democracy.
- A Conceptual Model to Guide AI Risk Governance Strategies (Mulligan + Marda + Wang, Knight Columbia, March 16 2026) (Deirdre K. Mulligan et al., Sociotechnical Risk Governance, March 16 2026) is a complementary framework that adds the harm-versus-hazard distinction and a sociotechnical-system orientation intended to give the Narayanan-Kapoor "resilience" framing operational content.
Sources
- AI as Normal Technology (Narayanan & Kapoor, Knight Columbia, 2025) — canonical
- AI Snake Oil (Narayanan & Kapoor, 2024) — book-length popular-press articulation of the same frame
- Digitalist Papers — Cochrane's "Just Relax" — closely-aligned rejection of preemptive AI regulation on different (economic-conservative) grounds
- Why I Think AI Take-Off Is Relatively Slow (Cowen) — compatible slow-takeoff framing from Tyler Cowen
- Anticipatory AI Ethics (Lazar, Knight Columbia, May 1 2026), AI as Social Technology (Farrell + Shalizi, Knight Columbia, May 11 2026), A Conceptual Model to Guide AI Risk Governance Strategies (Mulligan + Marda + Wang, Knight Columbia, March 16 2026) — Knight Columbia 2026 follow-ons
See also
- Law of Accelerating Returns (LOAR) — the framing AI as normal technology most directly opposes
- San Francisco Consensus — the set of beliefs AI as normal technology rejects
- Compressed 21st Century — Amodei's version of the acceleration claim
- Containment (Suleyman) — described as over-designed for a threat profile the framework holds will not materialize