Arvind Narayanan at ICML 2026: AI is 'normal technology'

A Princeton computer scientist stood in front of the ICML 2026 audience in Seoul last week and told a room full of AI researchers that their own field is not, on any evidence he can see, about to end their jobs. Arvind Narayanan's keynote, 'What will be left for us to work on?', is worth reading precisely because he refused the two easy answers, that everything is fine, or that nothing will be left.

His argument runs in three parts. The 'AI as Normal Technology' framework, he told the audience, is correct and useful as a way to think about AI's impacts, unless and until there is some future discontinuity such as through recursive self-improvement. Recursive self-improvement, he added, is worth taking seriously, but there is no milestone that companies might achieve in the lab that will suddenly put us all out of work. And the future of work is real, but it depends on humans developing what he calls complementary skills, meaning judgment and taste, evaluation of model output, setting quality standards, and deciding where AI should and should not be deployed.

The empirical anchor of the talk is a gap his team has been measuring. Over 24 months, model capability has 'shot up dramatically', while reliability, meaning consistency, robustness, calibration, and whether failures are recoverable, 'only increased by five or ten percentage points'. That gap is why the impressive agent demos have not yet translated into wholesale role automation, and it is why he thinks the real work of the next decade is organizational and evaluative rather than another leap in model scale.

The historical shape of the argument will be familiar to anyone who has watched this debate before. ATMs did not empty out bank branches, radiology employment grew as AI made imaging cheaper, and translation remained stable despite near-human AI performance a decade ago. Narayanan's framing of the alternative he wants is 'co-superintelligence', a bet that 'AI-augmented humans' will outperform 'AI systems acting alone', provided humans keep control over deployment decisions.

The honest caveat is that this is one researcher's read, delivered on friendly turf for the 'normal technology' framing that Narayanan is associated with, and the talk does not tell you which specific benchmarks produced the five-to-ten-point reliability figure or how the argument extends beyond knowledge work. What it does give you is a concrete prescription, '10 hours per week learning and experimenting' with new workflows, and a warning against handing things to a black box. If he is right, the shift in what work is worth doing is already here, and it looks less like unemployment and more like a change in what the job itself actually is.