Carl Benedikt Frey, the Oxford associate professor of AI and work behind the 2013 Frey–Osborne estimate that 47% of US jobs were susceptible to computerisation, argues that technological progress is never guaranteed. It depends on decentralised, competitive systems that let people explore. LLMs, as he sees them, recombine rather than invent, and AI will only pay off broadly if it creates new industries instead of merely automating old ones. Meanwhile workers need protecting as people, not as jobholders.
Progress is not inevitable. In his new book, How Progress Ends, Frey says that if progress were inevitable, the industrial revolution would have come earlier and most countries would be rich. Song China, the Dutch Republic and Victorian Britain each led only briefly. Innovation has an exploration phase and a scaling phase. The Soviet Union scaled steel and heavy industry for decades, but its engineers had only a handful of funders. It collapsed as computing demanded exploration, and he argues geography, institutions and culture alone cannot explain that. In the US, Bessemer passing on Google in 1999 did not kill it.
Weak ties and the lobbying trap. Frey cites weak-tie networks as a "collective brain". Katalin Karikó found collaborators for mRNA work despite decades of rejection, and Page and Brin met at Stanford. What matters, he says, is the autonomy to choose collaborators, not democracy as such. Once process gains run out, incumbents protect themselves instead of innovating, as Kodak and Google did. He cites research that inventors are less productive inside incumbents than at their former startups. US antitrust against IBM and AT&T opened software and the internet to new firms, while Japan's keiretsu blocked entry and Japan stagnated after leading in the 1980s.
What LLMs cannot do. A model trained on data to 1900 would not have predicted flight. Frey says AI still awaits a "separate condenser" that would make it resilient to a changing world, and cites a 2023 result in which amateurs beat superhuman Go programs. He is impressed nonetheless: models out-tutor him. He sees trust, and trusted brands, growing in value.
Automation versus new work. Whether a job changes or vanishes matters little to a displaced bus driver, he says. If AI is used mostly for automation, growth will be short-lived. New sectors such as railways, cars and electrical goods created the jobs that came before, and a few one-person unicorns would not spread prosperity.
Policy and individuals. Frey expects distrust of driverless cars to fade, as it did for automatic elevators. White-collar work is tradable, so he expects it to move abroad as barriers fall. He favours Danish-style flexicurity and generous safety nets, noting that areas with more generous poor laws saw less Luddite-era unrest. For individuals, he points to in-person communication, adaptive creativity and manual trades. Answering a listener, he says societies once tolerated deadly early railways and exploding tractors but are now more cautious. That makes reliability vital, he argues, though over-regulating before deployment would be a mistake.