Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
In the San Francisco Bay Area, one often hears the same story about AI. The technology, according to many who work at the labs, will create a “permanent underclass”. Anthropic chief executive Dario Amodei referenced the phenomenon in an essay this year, estimating that 50 per cent of entry-level white-collar jobs will be disrupted within five years.
The idea is as follows: as the models approach artificial general intelligence, a small tech elite will own the AI platforms, data centres and energy that power much of the economy. Agentic AI and advanced robotics will replace labour as we know it and most jobs will disappear. The majority of the population, according to the San Francisco view, will have few assets and struggle to earn a decent income selling their labour.
Economists should be sceptical of grand narratives that lack hard evidence. Certainly, the views of tech employees are conjecture. Labour market data provides scant, inconclusive proof of AI disruption. And some of the most eminent economists have predicted that AI will have a more modest impact.
But we are in the early days of AI adoption. And the “underclass” theory, though perhaps hyperbolic, could gesture at some uncomfortable truths. If gains accrue primarily to capital holders and some expert employees, labour market dynamics could deskill other workers and push down the cost of their labour. If human capital investment in low-wage workers declines as a result, society could move towards a scenario that somewhat resembles the San Francisco view.
Research suggests that automation could drive bifurcated labour-market outcomes. A paper by David Autor and Neil Thompson finds it can either boost wages by replacing low-expertise tasks or decrease wages by replacing high-expertise tasks. Workers with knowhow that cannot be substituted by machines could see massive wage gains, while others could be deskilled and lose their livelihoods. A recent Anthropic study similarly demonstrated higher returns from Claude Code to high-expertise users. (Amodei has previously referenced Autor’s other work on skill-biased technological change.)
As AI boosts productivity, gains from innovation could concentrate in the hands of those with capital and expertise (that is, a form of human capital). A study on patents filed in the early 2000s provides some evidence for this, showing that for every dollar of value added by a new invention, 30 cents went to workers and 70 cents became profit. Most of the additional worker pay accrued to employees with already high salaries, who presumably had higher expertise.
We shouldn’t be surprised if economic gains from AI accrue primarily to those building the data centres and frontier models. Though economists tend to assume that the labour and capital shares of production stay constant in the long run, the San Francisco crowd insists that agentic AI will substitute capital for labour and shift this paradigm.
Second-order labour-market effects of AI could amplify the gap between high- and low-wage workers. As deskilled workers lose their jobs, an oversupply of newly deskilled labour could push down wages for all low-skilled workers. And while there will probably be new low-skilled jobs that arise from AI models’ tremendous demand for data, such as roles in data gathering and aggregation, it is hard to know today how they will look and pay.
Productivity divergence could have long-term effects. Upskilling and reskilling workers is difficult. Moreover, research suggests parental wage shocks can weigh on skills development in children, transmitting wage effects on to the next generation. Evidence suggests that workers with lower incomes invest in human capital at lower rates, blunting social mobility. As productivity between workers diverges, the incentive to invest in education and training for low-income people could fall, amplifying the bifurcation.
Some don’t buy the “underclass” view. Instead, they argue, AI productivity gains will boost production, push down prices and benefit consumers. But this depends on how the technology augments production: AI will create lots of cheap software, but will do less to boost the quantity of less supply-elastic essentials, such as housing. Unevenly distributed productivity gains could make these goods less accessible for deskilled workers.
Though a “permanent underclass” is not an imminent reality, in the near term AI could make some people richer and others worse off. Policymakers could today strengthen the social safety net in anticipation of bifurcated labour-market outcomes. But if AI changes the economy in fundamental ways, democratic societies might need to reimagine how productivity gains can better serve the public good, lest the Bay Area’s pessimistic visions materialise.
Send your thoughts to [email protected].
Food for thought
What do high minimum wages have to do with robots and automation? This paper explains.
Free Lunch on Sunday is edited by Harvey Nriapia
