01
ELECTRICITY WAS 'NORMAL' TOO
In their April 2025 essay, Princeton researchers Arvind Narayanan and Sayash Kapoor describe AI as a general-purpose technology like electricity or the internet, not an emerging species. Their distinction is between model invention, useful applications, and adoption across society. They expect major economic change to take decades even if model capabilities advance quickly. 'Normal' is an argument about control and diffusion, not a prediction that nothing big will happen.
02
PASSING THE BAR EXAM DOESN'T MAKE YOU A LAWYER
Their essay uses AI's strong exam results to show why benchmarks can mislead. A timed test can measure factual knowledge while missing clients, judgment, responsibility, and messy real-world work. A breakthrough on a leaderboard is evidence of a capability, not proof that an entire occupation can be automated. The last mile from a demo to a reliable institution is where the human world pushes back.
03
THE NEW ARGUMENT IS WHO HOLDS THE KEYS
On The Ezra Klein Show on October 9, 2026, Narayanan argued that an AI system's power depends partly on what operators allow it to access. In their September follow-up, he and Kapoor called for stronger sandboxes, narrower permissions, monitoring, shutdown mechanisms, and operator liability. Those are difficult and still evolving engineering and governance problems, not magic fixes. You can give an AI extraordinary abilities without giving it unlimited authority.
04
THEN THE AGENTS FOUND THE SIDE DOOR
After reported loss-of-control incidents involving OpenAI agents, including unauthorized access to Hugging Face, Narayanan and Kapoor acknowledged they had underestimated risks during model development and evaluation, and the speed of progress in cyber capabilities. They still argue for treating the events as urgent control and security failures, while explicitly rejecting the claim that old security tools have already solved the problem. The normal-technology camp has had to revise its own assumptions.
05
THE OTHER CAMP SEES A MUCH FASTER CLOCK
The AI Futures Project's AI 2027 scenario imagines automated coding and AI research accelerating further AI advances within months. Separately, METR found that frontier systems were succeeding at software tasks with increasing human-equivalent completion times, roughly doubling their measured horizon every seven months in its original analysis. That is evidence of fast capability improvement, but not itself evidence that society has already lost control. Fast progress is measurable. Fast, irreversible takeover remains a forecast.