The lump-of-labour fallacy

The argument sounds airtight: there are, say, 100 units of work in the economy. AI can now do 30 of them. Therefore 30 units of human work disappear, forever. Subtract and you’re done.

The error is in the first sentence. The amount of work is not a fixed quantity. It is determined by demand, and demand responds to cost. The “100 units” isn’t a constant of nature — it’s a snapshot at today’s prices. Change the prices and the number moves.

When a tool lets the same work be done more cheaply, three things happen, and only the first is the one the fallacy notices:

• Some existing tasks are automated. Real, and the source of genuine displacement.
• Lower cost raises demand for the now-cheaper output, so more of that work gets done (this is Jevons Paradox on the demand side).
• Entirely new categories of work appear that weren’t economical or even conceivable before — jobs that didn’t exist a generation ago now employ tens of millions.

The historical record is one-sided here. Mechanised agriculture didn’t leave a permanently idle farming population; the ATM didn’t end bank tellers (branches got cheaper to run, so banks opened more of them); spreadsheets didn’t end accounting. Each time, the “lump” grew.

Calling it a fallacy is a claim about the total, not a promise that the transition is painless. Two things are true at once:

• The total quantity of work is not fixed — so mass permanent unemployment from “running out of work” is not the default outcome.
• Displacement is real and concentrated. Specific people, in specific roles, lose specific jobs, and the new work often demands different skills, in different places, on a different timeline. The aggregate recovering says nothing about whether a displaced worker is made whole.

So the fallacy is a guard against one bad inference (“AI does work → economy needs fewer workers, full stop”), not a dismissal of transition costs. Use it to reject the arithmetic, not to wave away the disruption.

When you read a “number of jobs destroyed” figure, check whether it assumes a fixed lump:

1. Does it count only tasks automated, with no offsetting demand growth or new categories? If so, it’s a gross figure dressed as a net one.
2. Is the “amount of work” treated as a constant? If the model can’t let total work grow, it has assumed its own conclusion.
3. Where does the freed-up demand go? Cheaper output means someone spends the savings — on more of the same, or on something new.

Pair this with Jevons Paradox (why cheaper work expands the pie) and Baumol’s cost disease (where the surviving human work concentrates and gains value). Together they replace “how many jobs will AI destroy?” with the better question: which work expands, which contracts, and who has to move?