Cheaper thinking, more work

The dominant story runs in two versions that reach the same place. The doomer version: AI takes the jobs and there is nothing after that. The accelerationist version: AI takes the jobs and that is wonderful. Both assume the same mechanism, and the mechanism is the part I do not believe.

The assumption is that the amount of cognitive work to be done is roughly fixed, so if a machine does more of it, humans do less. Every serious automation wave has falsified that assumption, and software falsified it hardest of all.

The Jevons shape

When steam engines got more efficient, coal use went up, not down. When compilers made machine code unnecessary, the number of programmers exploded. When spreadsheets automated the work of an entire class of clerks, financial analysis did not shrink — it grew until it consumed more people than the clerks it replaced, because analysis that had been too expensive to bother with was suddenly worth doing.

The pattern is consistent: when the unit cost of something useful collapses, the range of situations where it is worth doing expands faster than the efficiency gain. AI is doing this to cognitive work right now.

There is an enormous backlog of thinking that nobody does today because it does not clear the cost bar. Every small business that never got a proper security review. Every dataset nobody analysed. Every legacy system nobody documented. Every contract nobody read closely. Not glamorous, individually small, and collectively vast. Drop the price by an order of magnitude and a lot of that becomes worth doing.

The bottlenecks that stay human

Cheap generation does not remove four things, and all four are jobs.

Specification. Someone has to say precisely what should happen, and that requires knowing the domain well enough to be precise. This is the scarcest skill in the entire stack and it is getting scarcer relative to demand.

Domain knowledge. A model knows what is written down. A lot of the important things are not written down: how this hospital actually schedules, why this factory line stops on Thursdays, which clause the counterparty always disputes.

Verification. As output volume rises, the value of being the person who can say “this is actually correct” rises with it. This applies from code review to medical diagnosis to accounting.

Accountability. When a system causes harm, a named human has to be answerable. You cannot outsource liability to a model. Any regulated field — finance, medicine, infrastructure, security — has this constraint at its centre, and it does not get automated away, because it is a legal and social structure rather than a technical one.

What I actually expect

Displacement, yes. Concentrated, painful, and unevenly distributed — junior roles that consisted mostly of producing volume are the most exposed, and I do not want to be glib about that. Someone whose job disappears does not feel better because the aggregate numbers are fine.

But displacement is not the same as elimination. My expectation for the next five to seven years is a productivity increase and a redistribution of what work looks like, not mass unemployment. The people who use AI as leverage get more done. The people who ignore it lose ground relative to peers, not to machines.

The failure mode I would actually worry about is not that there is no work. It is that the transition is fast enough to hurt people who did nothing wrong, and the mechanism for helping them lags by a decade. That is a policy problem, not a technology problem, and it deserves more attention than the timeline arguments get.

Why the loud version wins anyway

Both extremes are useful to their tellers. “AI takes all the jobs” is a great fundraising line if you are selling the AI, and a great engagement line if you are warning about it. “Nothing changes” is comfortable. The middle claim — significant disruption, significant new demand, painful transition, no apocalypse — is accurate and boring, and boring does not travel.

I am on the boring side. So far it keeps being right.