Princeton AI researcher argues AI’s economic impact will unfold over decades, not abruptly

Arvind Narayanan, a Princeton University computer science researcher, told an ICML 2026 keynote in Seoul that AI’s impact on work will arrive through a slow, multi-decade adaptation phase rather than a sudden discontinuity. He presented the ‘AI as Normal Technology’ framework, which compares AI’s trajectory to historical technologies like electricity and describes four phases: methods/capabilities, products/applications, early adoption, and structural adaptation. Narayanan emphasized that the adaptation phase has not yet begun even in early-adopter fields such as software engineering, and he argued it will unfold over decades.

Narayanan acknowledged that recursive self-improvement could alter this trajectory but said he does not expect lab milestones to trigger abrupt, economy-wide displacement. He contrasted two camps: one that treats AI as an imminent replacement for most human work, and another that sees AI as an amplifier of human potential. He urged the AI community to reject the replacement narrative to avoid provoking political backlash and to invest in complementary skills such as agency, taste, and judgment.

The keynote drew on Narayanan’s prior work with Sayash Kapoor on the ‘AI as Normal Technology’ framework, which he described as a causal model for how AI capabilities affect the economy and society. He noted that the framework is not a slogan but a 15,000-word essay being adapted into a book, and he cautioned against conflating ‘normal’ with ‘mundane,’ since AI remains a transformative technology on the scale of the industrial revolution.

Narayanan also discussed how AI tools like coding agents are evolving from experimental ‘vibe coding’ toward more sophisticated agentic engineering, but he stressed that even in software—an early-adopter domain—the adaptation phase has not yet arrived. He speculated that if future coding agents can reliably produce large, secure codebases, software may shift from mass-market products to extreme personalization tailored to individuals or teams, reflecting deeper structural change in the industry.

He concluded with a vision of human/AI ‘co-superintelligence’ and called on researchers to focus on evaluation methods that go beyond benchmark gains to assess real-world deployment factors. The talk framed the current moment as one of both excitement and anxiety within the AI community, arguing that proactive adaptation now can prevent later harms and missed opportunities.