Here is a question that sounds stupid until you sit with it for ten minutes.
A Question That Sounds Stupid
Why does a radiologist earn $400,000 a year?
The obvious answer is: because reading a scan is hard, and few people can do it. That answer has been quietly false for a while now. Machines have matched or beaten human radiologists on narrow detection tasks since roughly 2018. The obvious answer has been dying for seven years, and the salary went up.
The real answer is that the radiologist is not selling a reading. The radiologist is selling a signature. And a signature is not an act of perception - it is an act of absorption. When they sign, the legal, financial, and moral exposure of that scan moves off the hospital's balance sheet and onto a named human being who carries malpractice insurance, a license that can be revoked, and a reputation that can be destroyed.
That transfer is the product. It always was. The diagnosis was the delivery mechanism.
What AI Structurally Cannot Do
We have spent three years cataloguing what AI can't do yet. Long-horizon planning. Physical dexterity. Taste. Every item on that list has a countdown clock on it, and the clocks keep running out earlier than expected.
There is exactly one item with no clock: a model cannot be a defendant.
This isn't a capability gap. It's a category error. You cannot punish a system that has no continuity of self, no assets, no fear, and no future to be deprived of. You can fine the company that deployed it - but that is not blame, that is a cost of goods sold, and firms are extraordinarily good at pricing those in. Blame, the real kind, requires a target that experiences the consequence. That target has to be a person.
So watch what happens to the price of things.
The answer goes to zero. The signature does not. And when the cost of production collapses while the cost of a required complement holds firm, the complement is where all the value goes. This is the most boring result in economics and nobody has applied it to the professions.
"Intelligence is becoming cheap. Culpability is becoming the scarce input."
How a Blame Market Forms
Markets for scarce inputs don't announce themselves. They show up as weird job listings and strange insurance products first. Here is what I think the next five years produce, in rough order.
1. The job description inverts. Postings stop reading "diagnose patients / draft contracts / audit accounts" and start reading "review and approve AI-generated output." This is already happening and is being universally misread as deskilling. It is not deskilling. It is the job being stripped down to the only part that was ever load-bearing. The reviewer is not there to catch errors - audit after audit shows human reviewers catch a pathetic fraction of machine errors, because approving is a fundamentally different cognitive act than producing. The reviewer is there to be the name on it.
2. Compensation decouples from skill and re-couples to exposure. Two professionals with identical output quality will be paid differently based on how much liability their signature carries. Pay tracks the size of the downside you personally eat. Which means - and this is genuinely strange - the most valuable professional in an organisation may be the one who is worst at negotiating their own indemnification.
3. Sign-off becomes an API. Someone will sell certified human approval as a service. A pool of licensed professionals, an SLA, a per-signature price, an insurance wrapper. It will be pitched as "compliance infrastructure." It will function as a wholesale market in human culpability, and it will be a very good business.
4. A spot price for a human name appears. Once sign-off is a service, it's quoted. The market price of a licensed human signature on a category of decision becomes a real, observable number - probably first inside professional indemnity underwriting, then publicly. When that number exists, everything downstream of it gets weird, because the price of blame is now a thing you can hedge, arbitrage, and speculate on.
5. Liability arbitrage - blame outsources like manufacturing did. This is the part I find genuinely dark, and I have not seen anyone write it down. Blame is cheapest where legal recourse is weakest. If your sign-off can be performed by a credentialed human in a jurisdiction where suing them is impractical, the economically optimal move is to relocate your culpability there. Manufacturing went to where labour was cheap. Accountability will go to where consequences are cheap. And unlike a factory, you can move it in an afternoon.
6. Professions lobby for mandatory human sign-off - and it will look like safety. Medical boards, bar associations, engineering bodies will push hard for laws requiring a licensed human in the loop. Every argument they make will be a safety argument, and many will be sincere. But the structural function of those laws is rent extraction on a scarce input, which is what licensing has always partly been. The novelty is that this time the rent isn't on the skill - the skill is commoditised - it's purely on the exposure. We are about to watch an entire generation of regulation get written to protect a resource nobody will name out loud.
The Uncomfortable Part
If this is right, a lot of career advice is now pointed at the wrong target.
"Become excellent, become irreplaceable" assumed excellence was the scarce thing. In a blame market, your moat is not your skill - it is your insurability. Your bankable asset is a long, auditable, boring track record of having signed things that did not blow up. That is a fundamentally different asset than expertise. You build it differently. You can have world-class judgment and be uninsurable. You can be mediocre and be extraordinarily insurable.
And there is a trap in it. The rational individual move in a blame market is to accumulate signature authority - it's where the money is. The rational systemic outcome is a class of highly-paid people who bear enormous exposure for output they did not produce and cannot fully verify. That is not a stable arrangement. It is a machine for manufacturing scapegoats with excellent salaries.
I don't think this ends with humans being replaced. I think it ends with humans being retained specifically as the part of the system that can be punished - and being paid extremely well for it, right up until the first mass-casualty case where the signature turns out to have been a formality. Then the whole thing gets repriced in a week.
Predictions, So This Can Be Wrong
Falsifiable, or it isn't an argument:
By 2028, at least one venture-backed company sells human sign-off as an API with an insurance product attached.
By 2029, professional indemnity premiums in at least one field decouple visibly from individual error rates and start tracking volume of AI output approved instead.
By 2030, a major legal case turns on whether a human reviewer's approval was "meaningful," and the discovery process reveals average review times measured in seconds.
Right now, watch job postings. The ratio of "produce X" to "review and approve X" language in professional roles is a live indicator, and it is already moving.
If none of that happens, I'm wrong and blame was never the product.
The One Line
We spent a century believing we paid experts for their expertise. We paid them for their exposure. Now that intelligence is nearly free, we're going to find out exactly what a human name was worth - because for the first time, it's the only thing left on the invoice.
Part one of three. Part two asks what happens now that a thought has a unit price. Part three asks where the next generation of experts is supposed to come from.
