The people training their own replacements

On 20 May, Meta laid off roughly 8,000 employees. In the same fortnight, more than 1,500 of the workers who survived signed a petition asking the company to stop recording their keystrokes, mouse movements, and screen activity to train its AI models. Meta's Chief Technology Officer, Andrew Bosworth, told staff in an internal note that there is no opt-out from the programme, which the company calls the Model Capability Initiative.¹

I want to be precise about what this is, because the ordinary framings will not hold it. It is not a story about layoffs, though there are layoffs in it. It is not a story about surveillance, though there is surveillance in it. It is a story about a company that has organised the labour of the people it employs so that their daily work produces the capability that will, on the current trajectory, reduce the headcount of the cohorts that come after them. The 72,000 who remain are the raw material for the system that will determine how many of them are still there next year.

The arithmetic is no longer hidden. Meta's capital expenditure guidance for 2026 is $125–145 billion, up from $37–40 billion the year before. The 8,000 layoffs were explicitly cited by company leadership as a mechanism to offset that bill.² On the Q1 earnings call, CFO Susan Li said, of headcount planning, "I don't know what ideal headcount looks like anymore."²

I have been thinking about Li's sentence for a week. It is, in its way, the most honest thing a Big Tech executive has said this year. Read it carefully. She is not saying she does not know whether to hire more people, or fewer. She is saying that the concept of an ideal headcount, the number of humans the business needs to do the work the business does, has lost its meaning for her. That is a description of a company in which workforce size is no longer set by the work to be done. It is set by whatever is left after the AI budget is paid.

This is what post-substitution workforce planning looks like when it is named out loud. Headcount is a residual variable now. You buy the capability ceiling first, and then you see how many people you still need around it. Meta is not the only company doing this — Microsoft cut 6,000 in January, Google has cut several thousand across early 2026, AWS restructured in the fourth quarter of last year — but Meta is the first to make the substitution legible on an earnings call, and the first to industrialise the loop by which the surviving workers train the capability that displaces the next cohort.

Who pays, and who has consented to pay. The Model Capability Initiative records behavioural data, keystrokes, cursor movements, what is on screen, from Meta employees as they work. Bosworth's note is unambiguous: there is no opt-out. The data is being collected to improve Meta's internal AI systems, the same systems being positioned to take on work currently done by humans inside the company.

You can defend this on the narrow ground that the devices are employer-owned and the employment is at-will in California. That defence does not address what is actually happening. A worker whose keystrokes are being recorded to train the model that will reduce demand for her keystrokes is not in the same position as a worker whose employer monitors productivity to improve scheduling. The data she produces is not being used to manage her. It is being used to obsolete her function. There is no version of "informed consent" that survives Bosworth's "no opt-out" — informed consent without the option to decline is not consent, it is notification.

I do not think this survives contact with GDPR in the way Meta seems to assume. The UK ICO's guidance on workplace monitoring requires a documented lawful basis, and "legitimate interest" is hard to sustain when the processing purpose is the development of systems whose deployment will materially affect the data subjects' employment. The EU AI Act's provisions on AI used in employment contexts add another layer. Meta's UK and EU workforce is not California; the legal exposure on the current design is real, and not yet tested. That will come.

The petition is a signal, not a constraint. 1,500 signatures inside a company of 72,000 is about 2.1% of headcount. The United Tech and Allied Workers in the UK and the Alphabet Workers Union have both cited the Meta case in public statements, but neither has a bargaining relationship with Meta. The petitioners have no recognised union, no collective contract, no formal mechanism to compel the company to change the programme. They have documented their dissent. That is what they have.

I want to be careful here, because there is a temptation, when writing about labour and AI, to either inflate the petition into the start of a movement or dismiss it as a venting exercise that lets management absorb the dissent and continue. Both readings are too neat. What 1,500 signatures actually does is establish a public record that the workers most directly affected by the programme did not agree to it. That record matters legally, it complicates Meta's ability to argue implied consent, and it matters politically, because it punctures the frame in which AI deployment is something workers and employers are doing together. They are not. At Meta, in May 2026, one party is doing it to the other, and the other has said so in writing.

The data she produces is not being used to manage her. It is being used to obsolete her function.

Whether this becomes durable organising is a separate question, and I do not think it has been answered yet. The conditions for tech worker organising have been weak for a long time — high compensation, individualised career structures, no labour-law architecture aimed at the sector, and a workforce that has historically identified with management rather than against it. AI deployment of the kind Meta is running may change some of that, because it makes the adversarial structure of the employment relationship visible in a way that pre-AI Big Tech mostly managed to obscure. When the company tells you, in writing, that your keystrokes are being recorded to train the system that will replace your colleagues, the old story about being on the same team is harder to keep telling.

The reassignment cohort matters, and does not change the structure. Meta moved roughly 7,000 employees from cut roles to a new internal AI unit nicknamed Draft.³ This is real job preservation, and I do not want to flatten it. For those 7,000, the transition is from one Meta role to another, probably at similar compensation, possibly at higher career trajectory. That is not nothing.

It also does not change the underlying substitution. The 7,000 reassigned are not net new jobs; they are reorganised existing ones, oriented around building the systems that the capex is paying for. Future Meta employment is being designed around a smaller core of people who build, train, and supervise AI systems, with the long tail of operational, support, and middle-management roles thinning. The reassignment is a transition mechanism, not a counterargument. Five years of this and the question is not whether Meta employs fewer people. It is whether the people it still employs are doing recognisably similar work to the people it employed in 2024.

What the productivity dividend argument misses. A serious counter-case exists. Some labour economists argue that previous tech investment cycles, the cloud shift of the early 2010s is the usual example, eliminated specific roles and produced net employment growth over the following decade as the new infrastructure enabled work that had not previously been economic.⁴ Whether the current AI cycle follows that pattern is genuinely contested. I do not think anyone honest claims to know.

But the productivity dividend argument, even on its strongest form, addresses aggregate employment over time. It does not address distribution, and it does not address consent. The workers Meta laid off in May 2026 do not get to wait out the decade in which the dividend might materialise. The workers being recorded by the Model Capability Initiative are not being asked whether they want their labour to fund a transition whose benefits will accrue, on the optimistic case, to a different cohort. The "in the long run" framing absorbs the specific people who pay in the short run and does not give them back.

This is the move I want readers to notice and resist. When the consequences of a deployment fall on identifiable people now and the benefits accrue to a diffuse population later, "net positive over time" is doing work that the people paying the short-run cost did not authorise it to do. Distributional incidence is the question. Aggregate dividends do not answer it.

What to watch. Three things, over the next twelve months. First, whether the petition turns into anything with structural form — recognised representation, a formal complaint to UK or EU regulators, a coordinated action across more than one company. Second, whether the ICO or any EU data protection authority opens an inquiry into the Model Capability Initiative; I would be surprised if none does. Third, whether other large employers, and not only in tech, adopt comparable behavioural-training programmes once Meta has demonstrated that the legal architecture in the US is permissive enough to allow it.

I want to end on what Susan Li said, because I think she has told us where this is going more clearly than anyone else. The CFO of one of the largest employers in the world does not know what ideal headcount looks like anymore. That is not confusion. That is a description of a workforce whose size is now downstream of a machine purchase, made by people who do not work there, on behalf of capability that the workers themselves are involuntarily helping to build.

The honest reading is that the people most affected by this transition were not asked, are not being asked, and on Meta's current design will not be asked. That is the consequence worth naming. The benefits of AI deployment are real. The harms are also real. What is happening at Meta is that one party has decided who pays for both, and the other party has signed a petition saying so.

Footnotes

¶