the blessing
and the toll

There is a particular word that gives the arrangement away. You submit. You submit a pull request, a design for review, a manuscript, a claim. The word predates all of these uses, and yet it has kept its oldest meaning intact; to submit is to place oneself under. Somewhere upstream of the thing you have made sits a person or a body whose function is not to make anything but to permit it, and the vocabulary we have built around that person is priestly and sovereign at once. Approval. Sign-off. Clearance. The imprimatur — literally, let it be printed. You do the work; someone else decides whether the work is allowed to become real.

Whoever that someone is, they are rarely thanked and often resented, and the resentment survives even a gatekeeper who is careful and right on the merits. A review team can be responsive and fair and correct, and the people waiting on it will still feel it as the place where velocity goes to wait. The same charge builds around a compliance desk, a licensing board, a planning inspector, a peer reviewer, an editor. What these have in common is not the kind of work they oversee, and not the temperament of the people who staff them. It is a shared arrangement: someone can do the work, has the competence to do it, and still cannot proceed until a second party, who did not do it, gives permission.

Not every gate of this kind is worth arguing about, and it helps to say early which ones are not. When a landlord approves a subletter, the approval does not turn on whether the subletter is any good. The landlord owns the thing being acted upon, and ownership, not expertise, is what the gate protects. Gates like that — grounded in property, in consent, in the plain right to decide what happens to what is yours — are mostly untouched by any shift in what machines can do, and they are not the subject here. The gate worth examining is the other kind, the one that justifies itself by knowing better: the senior who signs off because they understand more, the credential that vouches for its holder to people who have no other way to know. These rest on a claim about competence, and a claim about competence is exactly what artificial intelligence has begun to unsettle.

You have probably seen the worry stated plainly, more than once. The machines will do the work, and the people who did it will have nothing left to do. That fear is real and near at hand, and nothing here is meant to wave it off. But it tends to arrive with a reassurance close behind — that the work does not disappear so much as move up a level, and that the displaced will not sit idle but become the ones who verify, approve, orchestrate, and evaluate what the machines produce. As a guess about where the jobs go, this is reasonable enough. What it may be slipping past, when we reach for it too quickly, is everything it takes for granted: that a society of verifiers is a fine place to live and work. That is the assumption worth slowing down on — not because displacement matters less, but because the consolation we offer against it may reshape work more deeply than the displacement itself, and we are accepting it almost without looking.

Two things we usually picture as one can, in fact, be separated. There is the making of a thing — the diagnosis, the drawing, the code, the argument. And there is the authorizing of it: the act that lets the thing count, circulate, and bind. For most of history the same hands did both, so we had no reason to see them as two things at all. The medieval guilds are the plain case. A guild was an association of craftsmen — weavers, masons, goldsmiths — that decided who was allowed to practice a trade and what the work had to meet. The master's mark stamped on a finished piece did two jobs at once, inseparably, because one person did them both: it told a buyer the work was sound, and it kept out anyone beyond the guild who might have competed.

Sociologists have a name for the second job. They call it professional closure, which is only the tendency of a trade to protect its members by controlling who is let in. Their recurring finding is how hard closure and quality are to tell apart, because the two almost always come bundled. When the United States remade medical training after the Flexner Report of 1910, it raised the floor of what a doctor had to know, which was a real gain — but the same reform shut down the schools that had trained poorer, rural, and Black physicians, which thinned the profession and lifted the incomes of those who stayed. One act did both, and there is no clean way to keep the good it did apart from the harm.

What kept the two bundled — what kept the resentment from boiling over — is that the person at the gate could usually do the work too. The master could make the shoe. The senior physician knew more medicine than the resident. The editor could, in principle, write the piece. Authorization rested on top of competence, and because it did, you could never quite say whether the fee you paid bought the assurance or the permission. That blurring flattered everyone it touched. It let the arrangement look like protection even where it worked as a toll, and it let the people collecting the toll believe the gentler version of what they were doing.

This is the bundle the machine pulls apart. Once a system can produce the work as well as the person who used to hold the gate, or better, the competence is coming from somewhere else, and the gate is left standing without the thing that held it up. The gate does not fall. What falls away is the justification — the quiet belief that the gate was doing more than gatekeeping. And once that belief is gone, we are left holding an itemized bill, asked for the first time to say what the permission was ever worth on its own.

The clearest place to read that bill is a case that stands for many others. Suppose a system can read a chest radiograph and produce a report as accurately as a board-certified radiologist — not in a staged demonstration but in ordinary practice, which is roughly where the field has been heading, and roughly why radiologists keep appearing on the lists of jobs said to be at risk. Suppose that before the report can enter a medical record and guide treatment, a physician still has to put a name to it, and that the signature costs a hundred dollars. The reading itself came from the machine. Whatever the hundred dollars is buying, it is no longer buying the read.

So what is left for it to buy? The reading is done, and done well, which means we are paying for something other than the reading, and the honest thing is to sit with what that could even be. What surfaces is not a tidy list so much as a braid, and the strands are hard to pull apart.

The sturdiest of them is accountability. A machine can produce a report but cannot be held to it; it cannot be sued, struck from a register, or made to carry the weight of a cancer it missed. Everything we have built for catastrophic error — malpractice law, insurance, professional discipline, the plain need of a grieving family for someone who will answer — rests on persons who can be named and punished and made to pay. The signature turns a probabilistic output into an accountable act, and the signer becomes the one who can be blamed. There is a case, and not a weak one, that a hundred dollars is cheap for a place to send recourse — for a human who will lose sleep, whom a family can look in the eye, who has staked a license on being right.

Wound through that is something quieter and harder to price: every so often the human catches what the machine got wrong. Most of the time the signature is a formality, and its whole worth hides in the one read in some hundreds where the model is confidently mistaken in a way a person would notice, or where the case falls outside anything the system was trained on. The value is real, but it is back-loaded, and that is exactly what feeds the resentment — to the person paying, the ninety-nine ordinary cases are all that show, and in each of them the gate looks like friction and nothing more. The catch that justifies the arrangement is, by its nature, the thing you almost never see.

And running alongside both, impossible to fully separate from them, is the toll: the chance that the signature is required not because it adds accountability or catches error but because a profession that controls the gate has reason to keep the signature mandatory, and the standing to keep it so long after it has stopped changing the read. None of this needs bad faith. A profession can sincerely believe its sign-off protects patients and at the same time be defending an income, and the sincerity settles nothing. The sharp edge is who pays. A mandatory hundred dollars on a read the machine did for free lands hardest on the people for whom a hundred dollars decides whether they get the read at all — and in a time when the gap between the richest and everyone else is about as wide as it has ever been, that is not a small thing to wave past.

The reason this resists a clean verdict is that the three sit inside the hundred dollars at once, in a proportion no one outside can measure. The person paying reads the whole of it as toll. The profession reads the whole of it as accountability. Both are motivated, and the truth is a ratio neither side can see and neither is much inclined to look for.

Notice, though, what the braid quietly takes for granted. It assumes the machine has done the work as well as the person signing, so that the only things left to pay for are responsibility, the occasional save, and rent. That holds where the work is mostly a matter of getting it right — is the shadow a tumor or not. A great deal of work is not like that. Ask what you are paying for when you send a design, a paragraph, a room, a menu past someone for approval, and the answer is seldom correctness; the machine can turn out a competent version of any of them. What you are really asking is whether the thing is any good — whether it has taste — and taste is not error-checking. It is judgment along dimensions that have no right answer, only better and worse, and a better and worse that belong to some particular sensibility rather than to the world.

That should slow us down, because it is a different situation from the radiograph. Where taste is the substance of the work, it is not clear the machine has done the work at all; it may have produced something correct and lifeless, and the person at the gate is not sitting on top of equivalent output but supplying the one thing that was wanted in the first place. The bundle never came apart. To call that person a gatekeeper misreads what they are doing — they are doing the work, only at the end of it rather than the start. And taste, when it is what you are buying, is the least legible standard there is: you cannot write it down, cannot prepare for it, cannot argue with it. You can only submit. Which may be why the gates that turn on taste — the design review, the editor's this isn't quite there — draw the most personal and least answerable resentment of all, even when the taste behind them is real and good.

Two quieter pressures sit under all of this, whichever kind of work we are in. Even when the machine's output is the better output, we discount it for being the machine's, and part of what the signature does is launder that output into something we will accept, because a person has vouched for it — and it is an open question how long we keep paying for that conversion once the discount stops feeling earned. There is also the plain fragility of the thing. The gate stands only as long as a law or a regulation or a shared agreement keeps holding it up, and a market leans on those supports without rest. Someone will offer the read, or the design, without the toll. Whether they are allowed to, whether people keep paying the premium once a cheaper path exists, whether the requirement survives its first serious challenger — these are political questions, not technical ones. The signature persists at the sufferance of an arrangement that capital has every reason to test.

None of this can be measured from the outside, but it can be interrogated. Put almost any gate to a handful of questions and the accountability begins to separate from the toll, and where a gate answers them tends to decide whether the resentment aimed at it is fair or misplaced.

Reversibility. When the decision cannot be undone and its downside is grave — a misread scan, a drawing that has to hold up a building — verification is doing real work, and the resentment of it, however sincere, is aimed at something that guards against a harm no one could take back. When the decision is cheap to reverse — a design change that rolls back next sprint, a line of copy fixed after it ships — the gate is mostly buying delay, and the objection has a point. Much of the fury at internal review comes, I suspect, from the gap between how solemn the gate is and how easily the thing behind it could have been undone.

Exposure. Does the verifier carry the consequence of their own blessing? The radiologist who signs a wrong read is exposed; the signature is a real assumption of risk. A reviewer who rejects a proposal often carries nothing — a wrongful no is invisible, costless, never traced back. That asymmetry looks to me like the engine of the worst animosity. Where the verifier holds the gate and eats the downside of a bad call, there is a discipline inside the arrangement; they are spending something when they act. Where the verifier holds the gate but bears no cost for how they use it, you have power without exposure, and people are quick to sense that even when they cannot name it. The felt injustice of a review gate is seldom that it says no. It is that saying no costs the one saying it nothing.

Asymmetry. Does the verifier actually know something the doer does not? The senior clinician checking a trainee's read knows more medicine, and the gate doubles as teaching. The functionary enforcing a rule the doer already understands knows no more than the person waiting, and the gate is pure position. The most acute resentment comes from the verifier who knows less than the doer and holds the gate anyway — which is precisely the seat artificial intelligence threatens to put the human verifier in: signing off on work they could not have produced as well, holding a gate whose competence now lives on the far side of it.

Legibility. Can the doer tell in advance whether they will pass? Where the standard is written down and stable, the gate behaves like a checklist, is only mildly resented, and — not by accident — is the kind of gate that automates cleanly. Where the criteria are matters of taste that cannot quite be stated, the gate feels arbitrary and personal, and the resentment runs highest, because the doer cannot even prepare. This is the ground where human judgment adds the most and where the human most resembles a tyrant, and the two are hard to separate for the same reason closure and quality were.

Run the examples through the axes and they sort. A structural engineer's stamp is irreversible, exposed, backed by real knowledge, and reasonably legible — which is why it outlived automation even where software could do the arithmetic, since what the building department buys was never the calculation but the assignment of responsibility to a licensed, insurable, punishable human. A cosmetology license required to braid hair — a real requirement in many states, and a favorite of the licensing critics — is low-stakes, unexposed, and rests on no knowledge the braider lacks; it lands near pure toll. Code review sits in the middle and turns on culture: healthy where the reviewer shares the fallout of a bad merge, corrosive where it is a ritual of status. Radiology before AI sat near the engineer's stamp. After AI it slides toward a stranger place, one we have no settled name for — and it does not slide alone. The same is coming for anyone whose role was to lay expert judgment over work a machine can now do: underwriters, translators, paralegals, auditors, the reviewing tier of half the professions.

This is where the abstractions touch a working life. If the machine does the read and the human supplies the signature, the wage has come loose from the skill: the physician trained for a decade to interpret images is paid now not for the interpretation but for being a legal person who can answer for it. Madeleine Clare Elish gave this arrangement a name from the study of automation — the moral crumple zone, where the human kept in the loop of a mostly automated system ends up absorbing the responsibility the system cannot hold, positioned less to prevent failure than to take the impact when it comes. The signature can become exactly that: a person held liable for a judgment they no longer meaningfully make, kept on so that someone remains to point at.

Whether that is a diminishment turns on one thing — whether the signature stays tied to real judgment on the hard cases, or decays into reflex, a stamp applied without looking because looking was priced out. The physician who reviews the machine's uncertain reads and overrides the wrong ones is doing something real, and being answerable for it is close to honorable. The physician who signs a thousand reads an hour without opening them is performing an accountability that no longer happens. The same seat holds both, and which one a person occupies is not always theirs to choose — which is worth holding onto before writing only from the side of the person paying. Whoever collects the toll did not build the machine that hollowed out their expertise, and often the gate is the last thing between them and no livelihood at all. You can think a toll unjustified and still owe its holder that much: the case against the gate and the sympathy for the person standing in it do not cancel.

There is a thread of hope under this, and it runs through economics rather than sentiment. The moment a human steps in gets counted as pure cost, the expensive exception the machine could not handle. But the exception is exactly where the system meets something it has not seen, and the person's handling of it is often the only way that new thing gets found, and later taught back into the machine. Seen that way, the one who steps in is not the friction slowing the system down but the place it learns.

Widen the lens from the single signature to the whole economy, and the question the essay opened onto comes back. If the consolation holds, and the displaced really do become the ones who verify and approve and evaluate, then everything turns on whether that work stays tied to consequence and judgment or floats free as a toll. Tied to real accountability and real work on the hard cases, a society of verifiers is a decent enough division of labor — machines on the mechanical part, people on the part that needs someone to answer for it, and some people, still, in the older seat of deciding what should be made at all. Floated free — the judgment automated away, the consequence insured and diffused until it rests on no one, the gate still standing and still charging for passage — it becomes an economy of empty blessings, where the resentment everyone feels is not a mood but an accurate reading of extraction. Which of the two we get is not something the technology decides. It is a choice about what to do with the remainder, and we are making it now, mostly by not noticing that it is a choice.

That resentment is best treated as information, and like most information it is neither to be trusted whole nor thrown away. It is a detector, decently tuned, for gates that have come loose from their justification — for the verifier who knows no more than you, carries no exposure, governs a reversible decision, and charges for passage anyway. When it fires at that, it is right, and the answer is not to soothe it but to ask whether the gate should exist at all.

The same detector misfires, though, and it misfires in the way most worth guarding against: on the gates doing their quiet, unglamorous work, where the verifier is the one thing standing between cheap, abundant, unaccountable production and a harm no one could undo and no one would answer for. There the ninety-nine trivial blessings are all that shows, the whole thing reads as toll, and acting on the reading removes the last place responsibility lived.

The technology settles none of this. What it does is take away the blurring we sheltered in — the comfortable fusion of competence and authorization that let every gate present itself as protection. Once the machine can do the making, each gate has to say plainly what it is for — whether it holds real accountability, catches the errors that matter, supplies the taste or judgment the machine still lacks, or only collects a fee for a signature that no longer touches the work. Most gates, looked at honestly, are some mixture of these, and the task is not to tear them all down or defend them all but to read the mixture in each one: who carries the consequence of the blessing, and whether the hundred dollars buys a backstop or a toll.

None of this collapses into a rule; it has to be worked out one gate at a time, through the fog. But the version worth wanting is not vague, and it helps to name it, so we know what we are reaching for.

It looks like the reviewer who opens the file instead of stamping it, catches what the machine misread, and in the catching gives the machine the thing it did not know — so that the person is not the friction in the system but the place it learns. It looks like the work where taste or judgment is the whole substance, where the human never stopped being the maker and the machine only set the table. It looks like keeping, for people, the older seat of deciding what is worth making at all.

None of that is a consolation invented to soften the losses. It is a real and better division of labor, and it is within reach — though it is not what we drift into. It is what we would have to choose, gate by gate, by insisting each time that the blessing stay tied to the thing that made it worth having in the first place: someone who looks, and who can answer for what they see.