AI as normal technology
Capability is not power. What changes the world is not what a system can do but what organisations manage to absorb, and that runs into structural speed limits measured in decades, not months.
- Horizon
- 3–10 years
- Evidence
- Projected
- Consensus
- Medium
This is not the site’s decorative counterpoint: it is a projection with a method, and it is held here to the same standard as the other five. Arvind Narayanan and Sayash Kapoor start from a conceptual move —separating capability from power— and from a clarification that tends to get lost when they are cited: seeing AI as normal is not underestimating its impact [773]AI as Normal TechnologyView source ↗.
The mechanism is structural speed limits on diffusion. One: critical sectors still use statistical techniques decades old, which opens a lag measured in decades between capability and consequential deployment. Two: adaptation requires redesigning workflows and restructuring institutions. Three: much organisational knowledge is tacit and unwritten, and that limits parallel learning. Four: feedback loops with the real world do not accelerate the way self-play does, because safety puts a ceiling on the size of each iteration. From there comes the assumption that does all the work: that adoption is itself a filter, because it requires demonstrating appropriate behaviour in ever more consequential situations.
The signals in its favour are measurable and are on this site: business adoption still in the minority in an official probability survey [225]Business Trends and Outlook Survey — National estimates (National.xlsx)View source ↗, aggregate productivity with no visible take-off [306]International comparisons show AI effect on productivityView source ↗ and the absence of labour displacement at the scale of the whole economy, reported as the main finding by the same authors as the youth-employment figure [164]Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial IntelligenceView source ↗. If the macroeconomic explanation from Zanna Iscenko and Fabien Curto Millet holds, that is direct evidence in favour [566]Looking for the Ladder: Is AI Impacting Entry-Level Jobs?View source ↗.
Its weak point is one of scope, and they half-acknowledge it: the safety-through-adoption thesis does not cover whoever does not adopt through institutional channels. The 2025 cyber-espionage campaign did not pass through any organisation that would have filtered it [69]Disrupting the first reported AI-orchestrated cyber espionage campaign (anuncio)View source ↗. The argument is strong against economic acceleration and weak against deliberate misuse, which is precisely where there is observed evidence. It is also worth not reading them as reassuring: they classify catastrophic misalignment as a speculative risk, but admit epistemic uncertainty about whether that risk is close to zero.
Profile
- SpeedVery slow
- ReversibilityReversible
- ConcentrationMedium
Assumptions that must hold
That diffusion bottlenecks are structural rather than temporary: critical sectors still running on decades-old statistical techniques, unwritten tacit knowledge and slow organisational restructuring.
That capability does not translate into power without passing through institutions that mediate it.
That the adoption process is itself a safety filter, because it demands demonstrating appropriate behaviour in increasingly consequential situations. This is the assumption doing all the work.
That actors who do not adopt through institutional channels are the exception rather than the main route by which harm arrives.
What would refute it
That usage intensity stops being marginal: that the share of work hours mediated by AI rises steadily instead of staying in the low range the authors cite.
That aggregate productivity measurably takes off in national accounts and not just in narrow task experiments.
That consequential deployments bypassing the institutional filter stop being exceptions: if the main harm route passes through no adopting organisation, the safety filter filters nothing.
That the labour effect concentrated at the entry rung generalises to the whole career ladder.
Early signals
- AI adoption by US firms (Census BTOS)Observed
Slow diffusion measured by an official probability survey: the datum that best supports the projection, and also the one that would refute it if it accelerated.
- AI adoption measured by transactions (Ramp AI Index)Weak signal
It cuts against in the margin: in a panel of young firms, paying for AI is the majority. The gap with the official survey is the distance between a panel and a representative sample.
- Youth employment decline in AI-exposed occupationsWeak signal
Deliberately ambiguous: its authors' headline finding is that there is no aggregate displacement, and the secondary one is a large and widening gap in one group. Both readings sit on the same dashboard.
«Not observed» is not a clean bill of health: it means nobody has seen it yet, which is different from it not happening.
View as table
| Indicators | State | Note |
|---|---|---|
| AI adoption by US firms (Census BTOS) | Observed | Slow diffusion measured by an official probability survey: the datum that best supports the projection, and also the one that would refute it if it accelerated. |
| AI adoption measured by transactions (Ramp AI Index) | Weak signal | It cuts against in the margin: in a panel of young firms, paying for AI is the majority. The gap with the official survey is the distance between a panel and a representative sample. |
| Youth employment decline in AI-exposed occupations | Weak signal | Deliberately ambiguous: its authors' headline finding is that there is no aggregate displacement, and the secondary one is a large and widening gap in one group. Both readings sit on the same dashboard. |
See among the scenarios →Report a mistake in this entry →
Sources
- [773] AI as Normal Technology · Knight First Amendment Institute at Columbia University 2025
- [164] Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence · Stanford Digital Economy Lab 2026
- [566] Looking for the Ladder: Is AI Impacting Entry-Level Jobs? · Economic Innovation Group 2026
- [225] Business Trends and Outlook Survey — National estimates (National.xlsx) · U.S. Census Bureau 2026
- [69] Disrupting the first reported AI-orchestrated cyber espionage campaign (anuncio) · Anthropic 2025
- [927] Prioritization of Risks from Artificial Intelligence: A Delphi Study of 272 International Experts · The University of Queensland / MIT FutureTech 2026
- [306] International comparisons show AI effect on productivity · Federal Reserve Bank of Dallas 2026