The Layoffs Are Not the Story. The Door Is.

The dominant frame on AI and employment, as of this week, is that a wave of corporate layoffs at Amazon, Meta, Cloudflare and Block proves the displacement thesis. The Washington Post published a piece on 21 May arguing the opposite: layoff rates in the United States are running at roughly pre-pandemic norms, and most businesses that say they use AI also say their headcount has not moved. The piece has been picked up across wire outlets. The conclusion most readers will take away is reassuring: AI has not, in fact, displaced very many people yet.

I think that conclusion is wrong. Not because the BLS data is wrong, which it isn't, but because the BLS data is measuring the wrong thing. The story is not in the separations column. It is in the hiring column, and specifically in the row marked "first job."

Start with what the layoff numbers actually say. Challenger, Gray & Christmas put AI-attributed layoffs at about 54,800 in 2025, roughly 4.5% of total US layoffs that year. Procap Insights, tracking through March 2026, has the running figure at 27,645, or 0.017% of total US employment. Both numbers are, on any reasonable reading, small. The NBER executive survey cited in the Washington Post analysis finds that even firms reporting high AI adoption report limited realised employment impact. If you take these three data points at face value, the displacement narrative collapses. It is not happening. It is barely visible in the aggregate.

This is also roughly what Andrew Jassy's memo, stripped of its rhetorical packaging, actually describes. The line every outlet quoted was "reduce our total corporate workforce." That is a restructuring sentence. It is not an automation-displacement sentence. Amazon, Meta, Cloudflare and Block all had genuine non-AI reasons to cut: a post-pandemic overhire to unwind, a higher cost of capital biting into discretionary headcount, product re-orgs that would have happened in 2019 too. The AI gloss is good for the share price and, more cynically, good for the CEO who would rather investors believe the cuts came from a position of technological strength than from one of having got the hiring wrong in 2021. The HBR and NYT coverage of the Jassy memo treated it as the AI-displacement smoking gun. I think it is closer to the opposite: it is what AI-washing looks like when a CEO needs cover for an ordinary cost cycle.

So the optimist's case is, on its own terms, reasonably tight. Layoffs are not elevated. AI-attributed layoffs are a rounding error. Executives are overclaiming on automation in their public communications. The doom narrative is overcooked.

Here is what the optimist's case misses. The standard displacement model the BLS implicitly measures goes: AI automates a task, the firm decides the worker doing that task is no longer needed, the firm lays the worker off, the layoff rate rises. If you build your evidence base on this model, and the layoff rate does not rise, you conclude AI has not displaced anyone. The problem is that this is not how firms actually adjust to a labour-saving technology, and it never has been. They adjust through attrition. A team of twelve loses two people to other jobs over a year, and instead of backfilling, the manager tries the work with ten, finds that with the new tooling ten is enough, and the two roles are quietly retired from the org chart. No layoff is recorded. No separation appears in the BLS data. The team is one-sixth smaller. Multiply across a corporate function and you have meaningful workforce shrinkage that the separations statistics will never capture.

The hiring side is where this shows up. And the hiring side is where, as the Washington Post itself notes when read carefully, the door is closing. Hiring rates have softened across the white-collar economy. The softening is concentrated in early-career roles. Even analysts who reject every other displacement claim concede this one: junior hiring in AI-exposed occupations has weakened in a way that the macro cycle alone does not explain. Young software developers in particular have seen disproportionate employment declines in exposed roles while their more experienced colleagues have continued to gain. This is not a layoff story. It is a "we are not opening that role this quarter" story, repeated thousands of times across thousands of firms, and it produces a labour market in which the people already inside it are mostly fine and the people trying to get into it are mostly not.

The separations column is flat. The new-hire column for juniors is the one that is moving, and that is the one the displacement debate is not looking at.

The distributional picture sharpens the indictment. Brookings, in January, put 6.1 million US workers in the overlap of high AI exposure and low adaptive capacity. The composition of that group is the part of this story that ought to be doing more work in the public debate than it currently is.

These are concentrated in clerical and administrative roles, and clustered geographically away from the metro areas where alternative white-collar employment is densest. This is not a random draw. It is exposure to a labour-saving technology stacked on top of an existing distribution of economic vulnerability, in a configuration that the existing distribution would have predicted. The same group that was last to benefit from the previous wave of professional-services automation is first to be exposed to this one.

Two honest counter-arguments deserve a hearing here, because the Brookings number is being used (including by me, just now) to do a lot of work. The first is David Autor's: exposure estimates have a long history of overstating realised displacement because they price in automation of tasks and price out the creation of new ones. ATMs did not eliminate bank tellers; word processors did not eliminate the people who used to type. The composition of the work shifted, and the headcount mostly held. The 6.1M Brookings figure is a static exposure measure, not a dynamic forecast, and treating it as a body count is a category error.

The second is that the Brookings estimate is about adaptive capacity, which is a different question from whether the technology actually displaces. A worker with low measured adaptive capacity in a high-exposure role might keep the job, with the job changing around them, and be fine. Or might not. The number tells you who is in the blast radius. It does not tell you what the blast does.

Both points are correct. I take them seriously. But I think they understate two things specific to this transition that the historical analogies do not capture. One: the new-task creation in previous automation waves happened on a timescale of years to a decade, and the affected workers were largely still in the labour market when the new tasks appeared. The pace of capability rollout in current-generation AI is faster than that, and the new-task creation, if it comes, has to come faster too. Two: the adaptive capacity gap Brookings measures is not random — it is correlated with the prior distribution of education, geography and household constraint. Even if the historical pattern holds, and exposed occupations re-compose rather than disappear, the re-composition asks workers to do something they may not be positioned to do. The bank teller who became a personal banker had a pathway. The 56-year-old administrative assistant in a non-metro location whose job is being automated into a Copilot workflow may have one too, and may not.

Which brings me back to the hiring freeze, and to what I think is actually going on. The argument I would make, which the data supports more cleanly than either the pure displacement narrative or the pure "nothing is happening" narrative, is this. Firms are integrating AI into workflows at a pace that is faster than the aggregate displacement statistics show, but slower than the layoff-announcement rhetoric implies. They are doing it through three channels, in order of magnitude. First, attrition-without-backfill, which the BLS cannot see. Second, deferred or cancelled junior hiring, which appears in the hiring rate but not in the separations rate. Third, actual layoffs framed as AI-driven, which are real but small and often overstate the AI content of the underlying decision.

The aggregate workforce, as a result, looks roughly stable. The people inside it look roughly fine. The people trying to enter it, and the people in the specific clerical and administrative roles in the Brookings 6.1M, do not.

This has a few implications I think are worth being direct about.

The first is that the "adoption is the binding constraint" frame, which I generally hold and have written from before, cuts in a less reassuring direction here than it usually does. The optimistic version of that frame says firms are slow, so the disruption is slower and more manageable than the capability headlines suggest. The version the hiring data supports is that adoption is moving fast enough to suppress junior hiring across entire functions, but slow enough that nobody currently employed has noticed yet. That is a politically and economically combustible combination, because the people who would normally protest the change are the people not yet in the system to protest from.

The second is that the AI-washing problem is doing real damage to the policy conversation. When Jassy frames a restructuring as an AI efficiency win, and the press treats it as evidence of displacement, and the displacement narrative is then debunked by reference to BLS data, the public conclusion is that AI is not affecting employment. The conclusion the data actually supports, that AI is affecting employment, just not in the layoff column, gets lost. The CEO gets the share-price benefit of the AI story, the workforce gets the labour-market consequences of the AI story, and the policy debate gets neither. This is a bad equilibrium and the press could fix a meaningful part of it by being more disciplined about which statistic is being claimed and which one is actually moving.

The third is that the entry-level pathway question is the one that matters and the one that gets the least serious work. The standard apprenticeship model in professional services, software, finance, and law works because firms hire more juniors than they need, the juniors do work that is partly training and partly genuinely useful, and a subset eventually become seniors. AI is hollowing out the partly-useful-work middle of that model. The juniors are still needed for the training-themselves part, but the firm no longer has the work to give them to pay for it. Either the firms invent a new way to bring juniors in that does not depend on billable junior work, which professional services in particular has shown limited appetite for, or the pipeline thins, and in a decade we discover we have not produced the seniors. The hiring-freeze data is the first visible edge of this, and it is the part of the current debate I am least sanguine about.

I am willing to be wrong about all of this in the direction Autor would push. New tasks may yet appear, the 6.1M may re-compose into adjacent roles, the junior hiring softness may turn out to be the macro cycle and not the AI signal. I would update on that evidence when it arrives. What I am not willing to do is take the Washington Post piece, read it as "AI is not displacing workers, BLS proves it," and stop there. The BLS measures what it measures. It does not measure the role that was quietly never opened, the junior cohort that was quietly never hired, the clerical function that was quietly absorbed into a tool a manager bought on a corporate credit card. The displacement, to the extent it is happening, is happening in the spaces between the statistics.

The layoffs are not the story. The hiring freeze is the story. And the part of the hiring freeze that is most worth watching is the part where the door does not open for the people who needed it to.

Footnotes

¶