Carl Benedikt Frey: "Professionals are not prepared for the coming changes"

AI will have the same effect on white-collar workers as the 1970s de-industrialisation had on blue-collar workers. But economies can, and must, adapt.

Carl Benedikt Frey is professor for artificial intelligence and work at the Oxford Internet Institute. His seminal 2013 paper “The Future of Employment: How Susceptible Are Jobs to Computerization” arguably created a first wave of anxiety about technological unemployment due to AI. His subsequent book “The Technology Trap: Capital, Labor and Power in the Age of Automation” explains why some countries are slower to adopt new technologies than others, based on an extensive analysis of the economic and societal changes during the industrial era. In September 2025, Frey published a new book, “How Progress Ends: Technology, Innovation, and the Fate of Nations” which was shortlisted for the Financial Times Business Book of the Year Award.

In recent months, the debate on AI seems to have shifted from safety concerns to the very specific anxiety about AI-induced mass unemployment. To borrow from the AI safety debate: What is your p(doom) for the labor market? How likely is it that AI will lead to mass unemployment?

It depends on the time frame. Over the next year, or even ten years, the risk is low. Over the next thousand years, it’s obviously higher. But even in the very long run, I don’t think the labor share will fall to zero.

We’ve always created new kinds of work. Most people today work in jobs that didn’t exist in 1940. Many still prefer in-person services like yoga classes, even though you can watch an instructor online. So while I’m open to the idea that AI may eventually outperform us in almost any domain, I don’t think the labor share will go to zero, even though it will be significantly lower than today.

In the short run, generative AI will lower barriers to entry rather than fully automate work. It makes content creation easier, just like GPS plus Uber reshaped taxi services. Professional service jobs are tradable and can be done somewhere else. If AI reduces the productivity difference between accountants in New York and Manila, this work may shift abroad. Global employment may not fall, but workers in Frankfurt or New York may feel as if their jobs have been automated away.

So that’s not a precise probability of mass unemployment, but that’s how I see it.

You recently wrote in the New York Times that AI-induced unemployment won’t just affect white-collar workers like accountants, but could reshape entire service cities like New York and San Francisco. You pointed to Detroit and Boston as past examples of economic transformation. What lessons should policymakers draw from history?

First, Detroit and Pittsburgh were heavily dependent on single industries. New York, London, and the Bay Area are more diversified. But diversification might be a thin cushion when the disruptor is a general-purpose technology that cuts across sectors – from paralegals to teachers, marketers, traders, and consultants.

Also, a banker or software engineer supports far more local service jobs than a manufacturing worker. On average, a tech worker supports about five local jobs; a manufacturing worker supported about 1.6 jobs. That means that AI might lead to a significant hit to these local economies.

During the manufacturing era, some places adjusted, and some places did not. As a general rule, cities did better when they erected fewer barriers to the creation of new industries rather than focusing on protecting old industries. The Bay Area benefited from the absence of non-compete clauses. Engineers could leave Shockley Semiconductor to start Fairchild, and later Intel. In Michigan, non-compete clauses were reintroduced in 1985, and new job creation faded.

A key conclusion from that is: Support industrial renewal and the development of new forms of human capital. College towns adapted better historically, though in the AI era that may not be true to the same extent. My view is that AI models may be better tutors but not as good for writing and research because they show very few signs of novelty. Humans remain better at experimentation, developing new skills through training and continuous learning. In those domains, humans will still hold a comparative advantage for the foreseeable future. And obviously, cities are hubs that facilitate interaction, human creativity and experimentation.

In your book The Technology Trap, you argue that governments often hesitate to adopt transformative technologies out of fear of social unrest. As a general-purpose technology, it seems that AI has a lot of potential for disruption as well. Do you see new political conflict lines emerging – perhaps a “white-collar Luddite” movement or even political parties channeling AI anxiety?

Not quite yet, though there are signs, for example when Hollywood actors and screenwriters were striking together for the first time. But I don’t think that many white-collar professionals are prepared for the changes that are coming. Traditional career paths in consulting, law, or finance are likely to be significantly less reliable. And very few people are prepared for the loss of status that may result from that.

As professional service jobs increasingly move abroad, this means fewer job opportunities for people who live here. And these people don’t want to compete at Manila salaries. So yes, I do think that we’re going to see some form of a Luddite backlash, but it’s not yet on people’s mental radar.

The counter argument of jobs moving abroad is that with the support of AI support graduates can advance much faster – reaching senior levels in a shorter period of time. Do you see AI accelerating careers rather than eliminating them?

Most people coming out of university say “I learned more relevant skills in the first six months of my job than in my education prior. And firms invest quite a bit in your learning but the trade-off used to be that you perform some valuable work in the meantime.

With AI, firms start to question how valuable that work really is and it is likely we’ll see firms investing a lot less in training. At the same time, companies still need a talent pipeline.

But we now actually have first empirical evidence that graduate hiring is slowing down. And while this predates AI, it has been accelerated and may be made permanent by it.

In the past, economic transformations like the Industrial Revolution also reshaped society: general education, social security and workers’ rights were all consequences of the Industrial Revolution. What should policymakers do now to ensure today’s transformation is fair and equitable? Do we need a new social contract?

Much of the welfare state can be traced back to the Great Depression, which showed misfortune can hit anyone, not just the poor. AI could similarly create affinity between middle and lower-income groups.

We don’t need to reinvent the wheel. People in Scandinavian countries worry less about automation than Americans do [because of the welfare state], but there’s probably more to be done to support the transition and retraining.

The hard part is that we don’t know what skills to invest in. Ten years ago, it was all about coding which now looks like a terrible bet. Bottom-up initiatives like Sweden’s transition agreements, where people decide more or less by themselves in which skills to invest are going to be important. Given that firms are less likely to invest in training themselves, we probably also need some industry-level training initiatives or governments that step in to provide relevant training.

Skepticism is healthy, but what’s the alternative to re-skilling?

Some argue reskilling hasn’t been very effective. Why double down on it?

True, many programs show limited results. But we rarely track outcomes long enough, or account for foregone earnings. Skepticism is healthy, but what’s the alternative? If displaced workers can’t retrain into meaningful work, the consequences could be disastrous.

Even if not every program pays off, providing opportunities is better than leaving people stranded. Until we reach a post-scarcity society, reskilling remains essential.

Diffusion of new technologies often takes decades. But AI is moving extremely fast; is it therefore likely that we can see the impact of AI on the labor market in a shorter timeline?

Exponential improvements in technology don’t always translate into exponential growth. Progress often bottlenecks in human institutions – clinical trials, regulations, resistant organizations.

I do think AI will reshape organizations and labor markets, but it may resemble the ICT revolution: Big impacts, but not necessarily sustained exponential growth. Without new industries, the productivity boost may be a one-off. Historically, real progress came from doing new things, not just automating old ones. If AI spawns new industries, they’ll take decades to mature.

Key Take Aways

Generative AI removes barriers for entry: In the short run, generative AI will increase competition in professional services as white-collar work moves from high-cost locations to places with lower cost of living.

Skills are important, but we don’t know which ones: It is difficult to predict which skills will still be sought after in the AI economy. Therefore, workers themselves should decide where they want to focus on and governments will likely need to step up funding.

Progress means doing new things: If AI is not used to create new industries, the productivity boost may remain a one-off. As hubs for creativity and experimentation, cities play a key role in creating these new industries, but must also prepare accordingly.