Sixteen thousand a month, and the people we've decided not to count

The Goldman Sachs US Daily that landed this month puts a number on something that until recently was being argued about as if it were still hypothetical. Elsie Peng's note estimates that AI-driven automation is now displacing roughly sixteen thousand American workers every month, about a hundred and ninety thousand a year, and that the burden of that displacement is falling on a specific group of people: workers in their twenties, in entry-level white-collar roles, doing the kind of structured cognitive work that large language models turned out to be unexpectedly good at.¹

I want to take the number seriously, because the response to it has already begun to follow a depressingly familiar shape. The first move is to call it small. Sixteen thousand a month sounds modest against a US labour force of a hundred and sixty-eight million and a normal monthly churn that runs into the millions. The second move is to call it temporary, a transition, the kind of thing that always happens when a general-purpose technology arrives, and that always sorts itself out in the medium run. The third move, which usually arrives a paragraph later, is the one about how new technologies create more jobs than they destroy, with the obligatory nod to bank tellers and ATMs.

I don't think any of those moves survives contact with what Peng's note actually says.

Start with the size. Sixteen thousand a month is not a large number in aggregate terms, and Goldman is careful to say so. But the aggregate is the wrong frame, because the displacement is not aggregated. It is concentrated in a cohort, Gen Z workers in their first or second white-collar role, and in a set of occupations that share a structural feature: they were the places where ambitious young people without elite credentials used to enter the professional economy. Data entry. Paralegal work. Customer service of the kind that required judgement, not just script-reading. Junior copywriting. The first rung on the bookkeeping ladder. The first rung on the analyst ladder. These are not romantic jobs. They are the jobs that, for several decades, have done the work of converting a community college graduate or a state-school humanities major into someone with a salary, benefits, and a CV that opens the next door.

That ladder is what's being pulled up, and the 3.3 percentage-point wage gap Peng identifies between high-AI-exposure and low-AI-exposure occupations is the early signature of it.² Wages don't fall first; entry positions disappear first, and then the wages of the people still doing adjacent work begin to drift, because the bargaining position of anyone whose job a model can plausibly do has weakened. The wage gap is not the harm. It is the readout of the harm.

Now the temporariness argument. The bank-teller story is the one everyone reaches for, and it's worth being precise about why it doesn't transfer. ATMs displaced a specific task, cash dispensing, while leaving most of what tellers actually did (account opening, problem-solving, sales) intact and indeed expandable, because cheaper branches meant more branches. The displacement and the offsetting opportunity were in the same building, often involving the same workers. The transition was real and it was painful for the people who lost their jobs, but the structural logic produced a new equilibrium that absorbed the affected cohort within roughly a working generation.

The current displacement does not have that shape. The tasks being automated are not narrow sub-components of larger jobs that will expand once the sub-component is cheaper. They are, in many cases, the whole job. A junior paralegal whose document review is now done by a model is not freed up to do higher-value paralegal work, because the higher-value paralegal work is also being done by a model, or by a more senior paralegal whose own productivity has gone up four-fold. The "freed-up capacity" framing assumes the freed person has somewhere obvious to go. Increasingly, in the occupations Peng is tracking, they don't.

This is the part of the argument where someone usually says: yes, but new jobs we can't yet imagine will emerge, as they always have. I want to take this seriously, because it is true that we cannot fully imagine the occupational structure of 2040 from inside 2026, and it is also true that previous waves of technological change did, eventually, produce new categories of work. But "eventually" is doing an enormous amount of work in that sentence, and the people who absorb the cost of "eventually" are not the same people who absorb the benefits of it. The graduate who can't get a foothold in 2026 does not get retroactively employed in 2034 when the new occupational categories stabilise. They get a gap on the CV that follows them for the rest of their working life, and a cohort effect on lifetime earnings that the labour economics literature has been measuring since the 2008 recession and is consistent enough to be considered a fact.³

Who is paying, then. This is the question the Goldman note implicitly answers and that the surrounding commentary has been reluctant to make explicit. The cost of the current AI deployment is being paid, disproportionately, by people in their early twenties who entered the labour market in 2024 and 2025 expecting the kind of on-ramp that their older siblings had, and who are finding instead that the on-ramp has been narrowed. They are paying it in the form of longer job searches, lower starting salaries, more time in roles that don't build the CV, and, this is the part that doesn't show up in monthly displacement counts, in the loss of the developmental experience that those entry roles used to provide.

The displacement isn't only of jobs. It's of the structure by which a generation learns to do professional work.

Because here is what an entry-level paralegal job actually was, before it became a thing a model could mostly do. It was three years of being in the room while senior lawyers argued about strategy. It was learning what a good brief looks like by drafting bad ones and being told why they were bad. It was building the judgement that, ten years later, made you a competent mid-career professional. When the model does the drafting, the model also absorbs the developmental loop. The senior lawyer doesn't stop existing. But the pipeline by which junior people become senior people has been quietly disconnected, and we are only at the beginning of finding out what that does to the profession over a decade.

The same logic applies in journalism, in software, in design, in accounting, in marketing. The cohorts entering these fields now are entering a labour market in which the bottom rungs are being removed faster than the middle and upper rungs, and the standard response, "they'll just start higher up", assumes a developmental shortcut that doesn't actually exist. You cannot start at rung four if rungs one through three were where you learned to climb.

What gets called inevitable. I notice that the framing around all of this has shifted, over the last eighteen months, from "AI might displace some workers" to "AI is displacing workers and this is the cost of progress." The grammar of inevitability has settled in fast, and it is worth saying clearly that it is a choice of grammar, not a description of reality. There is nothing inevitable about deploying models in ways that eliminate entry-level positions rather than augment them. There is nothing inevitable about firms capturing the productivity gains as margin rather than passing them through as expanded hiring. There is nothing inevitable about a tax and transfer system that does not redistribute any of the gains to the cohort bearing the costs. These are decisions, made by specific actors, with specific incentives, that could be made differently.

I am not arguing they will be made differently. The political economy of the moment does not point that way, and I am not in the business of pretending otherwise. But the language of inevitability does a particular kind of work: it converts the decisions into weather, and weather is something you adapt to rather than contest. Sixteen thousand displacements a month is not weather. It is the cumulative outcome of a very large number of procurement decisions, deployment decisions, hiring freezes, and quiet attritions, each one made by someone who could have made it differently.

What I'd watch. Three things, over the next twelve months. First, whether the 3.3 point wage gap widens or stabilises, wage compression in adjacent occupations is the leading indicator of how far the displacement is propagating. Second, whether graduate hiring at the large professional services firms (law, accounting, consulting, the historic absorbers of educated entry-level labour) holds at 2024 levels or contracts further; the early signals from the 2025 graduate recruitment cycles were not reassuring. Third, whether any serious policy response materialises that treats the affected cohort as owed something, rather than as the unavoidable friction of a transition that, on current trends, will mostly benefit people who were already doing well before it began.

None of this means the technology shouldn't exist, or that the productivity gains aren't real, or that we should be trying to put the model back in the box. Those aren't the questions. The question is who bears the cost of the transition, and whether we are going to keep pretending that the answer is no one in particular.

The Goldman note suggests the answer is sixteen thousand people a month, mostly under thirty, and that we have so far decided not to count them as anyone we owe a response to.

Footnotes

¶