Princeton Professor Narayanan at ICML 2026 Keynote Calls AI Normal Technology

Princeton professor Arvind Narayanan delivered the keynote address at ICML 2026 in Seoul. He argued that AI should be treated as normal technology rather than an existential break. The speech laid out three points that shift attention from sudden job loss to steady adaptation.

Narayanan stated that laboratories will not produce a single milestone that leaves every worker unemployed. He added that recursive self improvement deserves study yet remains distant in practice. Listeners heard that work itself will change and that preparation must begin now.

The professor urged the AI community to reject the idea that tasks will simply be handed to machines. Instead he called for deliberate development of skills that complement current model limits. Judgment and taste were named as examples of abilities models still lack at scale.

Narayanan defined AI as normal technology by comparing it to earlier shifts such as electricity and computing. He said progress follows predictable patterns of adoption, cost reduction, and workflow adjustment. Sudden leaps that empty offices do not appear in the historical record.

The audience received data points from past automation episodes. Employment in affected sectors continued after initial drops because new roles emerged. Narayanan used those records to frame current fears as understandable but overstated.

The main tension in the speech sits between forecasts of total replacement and measured evidence of gradual change. Narayanan positioned himself against both alarmist claims and complacency. He pressed researchers to study real capability curves instead of dramatic scenarios.

He noted that claims of rapid takeoff rest on assumptions that remain untested. No present lab result shows models rewriting their own code at superhuman speed. The gap between current systems and those hypothetical models stays large.

One section examined the skills humans can still claim. Narayanan listed evaluation of model output, setting of quality standards, and selection among conflicting suggestions. These activities require context that models acquire only through human direction.

He gave the example of product decisions that rest on unclear user preferences. Taste developed through experience allows teams to choose among technically sound options. Models can generate candidates but cannot yet rank them against unstated goals.

Narayanan closed with a short vision he labeled co-superintelligence. The term describes combined human and model performance that exceeds either alone. He said the outcome depends on active skill building rather than passive use of tools.

The speech received measured applause. Attendees noted that the tone differed from previous keynotes focused on capability races. Several researchers said they plan to test the suggested skill categories in upcoming projects.

Future signals to watch include lab reports on self improvement benchmarks and hiring data for roles that emphasize judgment. Continued absence of sudden capability jumps would support the normal technology framing. Shifts in job postings that reward evaluation skills would offer early confirmation.

Narayanan reminded listeners that adaptation is an active task. Waiting for displacement provides no advantage. The work of building complementary abilities starts with the next project rather than the next breakthrough.