On the lump of labour fallacy

Marc Andreessen is right, but oh so wrong

The ATM was going, going, gone

There is a famous story in economics about how ATMs did not, in fact, automate the bank teller. James Bessen popularised it in his 2015 book Learning by Doing: despite the proliferation of cash machines from the 1970s onward, bank teller employment held steady or grew. ATMs lowered the cost of operating a branch, so banks opened more branches. Tellers shifted from counting cash to selling mortgages and managing relationships. In practice, technology complemented labour rather than displacing it. The story became, as Matt Yglesias put it, a ‘load-bearing parable’ about the myth of technological unemployment.

The problem, as David Oks recently pointed out, is that the story has a second act nobody mentions. Right around the time economists started celebrating how ATMs hadn’t killed the teller, the iPhone did. By contract, mobile banking didn’t automate the teller’s tasks. It made the teller irrelevant by creating an entirely different way of banking. Teller employment entered sustained decline in the 2010s and has not recovered.

The distinction matters. Is AI a technology closer to the ATM or iPhone, and how soon till we find out?

Don’t take my job

The last two weeks have been awash with tweets, articles, and news stories about AI induced unemployment.

Marc Andreessen came out swinging on the other side of the argument, arguing that AI-related layoffs are “all fake.” Companies are 25-75% overstaffed from the pandemic hiring binge, he says, and AI is the “silver bullet excuse” to clean house.

More formally expressed, he’s saying that anyone worried about AI destroying jobs is committing the lump of labour fallacy, the belief that there is a fixed amount of work in an economy. The term dates to 1891, when economist David Frederick Schloss coined it to rebut claims that shorter working hours would reduce unemployment. Schloss argued the amount of work is not fixed. New technology creates new demands, new industries, new jobs.

This also fits in neatly to the theory that humans have an unquenchable demand for goods and services. Satisfaction doesn’t last long. Consumer expectations (whether in term of quality, speed etc) can rise indefinitely.

It’s true the Luddites smashed looms and the textile industry still employed more people a generation later. Spreadsheets were supposed to kill accountants. ATMs were supposed to kill tellers (see above, with caveats).

Undoubtedly, Andreessen is onto something here.

Yet I worry he’s missing the broader point. Saying ‘she’ll be right in the long run’ may well be true... but we don’t live in the long run, we live in the here and now.

The body count

In practice, history teaches us some messy lessons about the lump of labour fallacy. When power spinning replaced hand-spinning in Britain, the transition was catastrophic for the people living through it. Hand-spinning employed roughly 8% of the population by 1770. Mechanisation began displacing spinners in the 1780s. Recent Oxford research documents the effects persisting until the mid-1830s. That is half a century of adjustment, during which entire communities were hollowed out.

The early evidence from AI follows a similar pattern at compressed timescales. Over 152,000 tech workers were laid off across 551 companies in 2024, and another 123,000 across 257 companies in 2025. About 20% of those layoffs were explicitly linked to AI and automation. Salesforce eliminated thousands of customer support roles. Fiverr cut 30% of staff while repositioning as “AI-first.” Block’s AI systems now handle 70-80% of customer enquiries without human intervention.

And the structural incentives point to acceleration, not stabilisation. A recent game theory paper from UPenn and Boston University, “The AI Layoff Trap” by Falk and Tsoukalas, formalises the problem. Each firm that automates captures the full cost saving but bears only a fraction (1/N, where N is the number of competitors) of the resulting demand destruction. This creates a prisoner’s dilemma: every firm rationally automates even when they collectively know restraint would raise profits. The authors tested six policy instruments (UBI, capital income taxes, wage adjustment, upskilling, worker equity, Coasian bargaining) and found that none of them alter the underlying incentive structure.

The retraining mirage

The standard policy response to every technological transition is ‘ increase retraining’. Even I’ve been fond of the line ‘AI will create more jobs than it destroys, the real roadblock is education.

Your role in a call centre may go away, but there’s plenty of demand for prompt engineers. Go out, pull yourself up by your bootstraps, and adjust.

This is a neat theory and sounds almost tantalising on first principles - but the evidence doesn’t support it.

Brookings published research in 2026 documenting the limits of worker retraining as a response to AI displacement. The evidence is sobering. Workers displaced by robotics adoption overwhelmingly ended up in lower-paid service jobs, not in the new technical roles the retraining was supposed to prepare them for. The World Economic Forum estimates 60% of workers will need retraining by 2027, but only half currently have access to adequate training. Interestingly, the United States chronically underfunds workforce development compared to peer nations - mainly due to the extreme costs of post school education.

There is an appealing irony in the idea that AI itself could solve this: personalised learning, adaptive curricula, training delivered at the speed of need.

It is plausible. It is also entirely untested at the scale required. “AI will retrain the workers AI displaced” is a proposition running more on optimism than evidence. I hope it turns out to be true, but building policy around an untested hypothesis while people are losing their jobs now is a particular kind of optimism.

The other angle here is geography. As job opportunities grow and die across regions - people will of course adjust. But adjustment takes time. Would-be migrants take months or years to recognise trends, find jobs, move families, get settled, then tell others their success stories. There’s a lot of pain and uncertainty in the meantime.

So who gets squeezed?

It’s becoming fairly obvious that the workers most exposed to AI displacement are not at the extremes of the skill distribution.

The highly skilled will embrace AI, augmenting their existing tasks with new-found speed and intelligence. Lawyers will use AI to draft contracts faster, economists will use it for data cleaning and sensitivity testing, doctors will use it for note-taking and diagnosis. High-touch physical roles (care workers, tradespeople, cleaners) are hard to automate for now because these jobs involve navigating unpredictable physical environments. The squeeze is likely in the middle (and squarely aimed at computer-based roles).

Acemoglu and Restrepo’s task-based framework describes the mechanism. Automation displaces workers from tasks they used to perform, reducing demand for their labour. New tasks are created, but there is no guarantee the displaced workers have the skills to fill them. The mismatch between displaced skills and new task requirements creates a wedge in the labour market that can persist for years.

Alex Imas, the Chicago economist, adds a demand-side point that sharpens the impact further on these roles ‘in the middle’. Even if AI doesn’t fully automate jobs, significant augmentation allows firms to hire fewer people to meet the same demand. Productivity rises. Prices may fall. But whether demand increases enough to reabsorb the displaced workers depends on elasticities that nobody can estimate with confidence.

The right fallacy, the wrong timescale

So where does that leave us?

Andreessen is right that the lump of labour is a fallacy. And it’s true folks make this argument about every whiz bang technology to come off the shelf. Technology does create more work than it destroys. Indeed it would be foolish to bet against it.

But the historical record is also clear on something else: the transition is brutal, it is unevenly distributed, and it takes far longer than the optimists assume. The bank tellers were fine until they weren’t, undone not by the technology everyone was watching but by one nobody was.

At its core the lump of labour fallacy is a statement about equilibrium. But it says nothing about the path to get there. For the graduate entering the workforce in 2026, the customer service agent whose role just got automated, the content writer whose clients switched to Claude, this uncharted path is the only thing that matters.

The interesting question was never whether AI would create new jobs. It of course will (and if you look at the current data, is).

The interesting question is how many people get stuck on the wrong side of the transition, where they live, for how long.

Invoking a 135-year-old term to explain why it will all be fine in the long run is fair enough - but it’s of little help to policy makers in the meantime.