Meta Is Watching Its Workers To Replace Them

On 20 April, Reuters reported that Meta has begun installing tracking software on its US employees' laptops.¹ The software, part of something Meta is calling the Model Capability Initiative, sitting inside a broader "Agent Transformation Accelerator" programme, captures mouse movements, clicks, keystrokes, and periodic screenshots. The purpose is not the usual one. Meta is not checking whether its engineers are at their desks or whether its recruiters are hitting activity quotas. It is generating training data. The captured work product, every drag, every paste, every Slack message composed and deleted, every form half-filled, is being fed into systems designed to build AI agents that can do the same work.

European employees are excluded from the programme. Meta's lawyers, reading the GDPR and the patchwork of national workplace-surveillance laws across the EU, concluded that what is legal to do to a worker in Menlo Park is not legal to do to a worker in Dublin or Munich. So the company does it in the place where it can.

I want to be precise about what this is, because the framing will matter. This is not a surveillance story, exactly, though surveillance is the mechanism. It is not a consent story, though consent is the thing missing. It is a story about a specific, almost clarifying arrangement: a company asking its workers to teach the software that will replace them, while recording them closely enough that they cannot plausibly refuse. It is, in a sense, the clearest version of the deal that a lot of AI deployment actually is. Meta has simply made it unusually literal.

This is not a surveillance story, exactly, though surveillance is the mechanism.

The mechanism

Start with what the software does. Keystroke and mouse capture, combined with periodic screenshots, is sometimes called "full telemetry" in the enterprise-monitoring trade. It is the most intrusive form of workplace surveillance in commercial use. The market for it, Teramind, ActivTrak, Hubstaff, Veriato, and a dozen others, grew rapidly during the remote-work expansion of 2020–2022 and has continued to grow since.² What is new at Meta is not the capability. It is the stated purpose.

Conventional employee monitoring is defended, when it is defended, as a tool of management: to measure productivity, to detect insider threats, to audit compliance. The worker is watched so that the worker can be managed. This is bad enough, and the evidence on whether it actually improves productivity is at best mixed and at worst negative.³ But it is at least a coherent story about why the watching exists.

Meta's programme is a different story. Here, the watching exists so that the work itself, the sequence of cognitive and motor operations that constitute the job, can be extracted, structured, and used as training data for a system that will eventually perform the work without the worker. The worker is not being monitored in order to be managed. The worker is being monitored in order to be modelled.

This distinction matters. Under traditional surveillance, the worker's continued employment is the point of the exercise, the employer wants more and better work from the person they are watching. Under this new arrangement, the worker's continued employment is, at best, incidental to the exercise, and at worst, the thing the exercise is trying to end. Every keystroke captured makes the keystroker marginally more replaceable. This is not a bug of the programme. It is the design.

Who knows, who consents, who pays

The Reuters reporting notes that internal backlash was immediate. That phrase does a lot of work. It implies that Meta did not, in advance, secure meaningful consent from the workers whose behaviour it intended to record and mine. It implies that the programme was announced as a fact rather than proposed as a question. This is consistent with a pattern I have watched accumulate across the AI-deployment landscape over the past two years: the people whose work is being used to train the systems that will change or end that work are systematically the last to be consulted.

We have seen this with writers whose published work was scraped to train large language models without licence or compensation, a fact now being litigated across multiple jurisdictions. We have seen it with artists whose portfolios were ingested into image models under the same conditions.⁴ We have seen it with customer-service workers whose recorded calls became training corpora for the chatbots that now handle the tier-one queue. In each case, the material that made the system possible came from workers whose consent was either not asked or asked in forms that did not permit a meaningful refusal.

Meta's programme is an intensification of this pattern, not a departure from it. What is new is that the extraction is happening in real time, on a live workforce, inside the employment relationship itself. The corpus is not a body of finished work collected retroactively. It is the ongoing behaviour of people currently at their desks. And because it is happening inside an employment relationship, the worker's ability to refuse, to opt out, to negotiate terms, to demand compensation for the use of their behavioural data, is structurally constrained in ways that a freelance illustrator being scraped was at least, in principle, not.

You cannot refuse a condition of employment and keep the employment. This is the core of what employment is. Meta's lawyers know this. It is why the programme is legal in California and not in Germany: the difference is not really about the technology, it is about how much a worker is permitted to refuse.

The geography of what's allowed

The EU carve-out is the part of this story I keep returning to, because it reveals the shape of the thing clearly.

Meta is not doing this in Europe because European workers have, through decades of organising, legislation, and case law, built a floor under what can be done to them at work without their structured consent. The relevant instruments include the GDPR, the Article 29 Working Party opinion on data processing at work, and a patchwork of national statutes: Germany's Betriebsverfassungsgesetz, which gives works councils co-determination rights over monitoring systems; France's labour code provisions on employer surveillance; Italy's Workers' Statute. These laws did not emerge from abstract principle. They emerged from fights. They emerged from workers who, in earlier technological transitions, refused to accept that new tools gave employers unlimited new rights.

The workers in Menlo Park do not have this floor. They have at-will employment, weak federal privacy law, and a company culture in which the framing "we are doing this because we can" is not, internally, considered an embarrassing thing to say. And so Meta is doing it.

What this tells us, what it has been telling us, loudly, for a decade, is that the question of how AI is deployed against workers is not a technological question. It is a question about the relative power of workers and employers in a given jurisdiction. The same company, with the same software, making the same business case, acts differently in Frankfurt and Fremont. The technology is identical. The power arrangement is not.

I think this is worth sitting with, because the dominant industry framing holds that AI deployment is a function of capability, what can the models do, what tasks can be automated, where is the frontier. The Meta programme makes it clear that deployment is at least as much a function of permission, what can the company get away with, where, and against whom. Capability is necessary. It is nowhere near sufficient. Between capability and deployment sits law, organised labour, consumer sentiment, and regulatory risk, and these variables vary enormously across jurisdictions. The map of where AI is being deployed aggressively against workers is not the map of where the technology is best. It is the map of where the floor is lowest.

The specific cruelty of the design

I want to name something about the Meta programme that I find, on reflection, genuinely striking.

Most automation throughout history has proceeded by observing work and then engineering a replacement for it. The Jacquard loom did not ask weavers to operate the loom while it learned to weave. The spreadsheet did not run alongside the bookkeeper, recording her every entry, so that she could be dispensed with. The observation was done at arm's length, by engineers and managers, and the resulting technology was then brought in.

Meta's approach is different, and more intimate. The worker is the observation instrument. The worker's own actions, captured in detail sufficient for imitation learning, are what make the replacement possible. There is no arm's length. The employee is both the subject being studied and, unwittingly or unwillingly, the researcher producing the data. The productivity of the worker in their job is precisely what degrades their position in the labour market over time, because the better they do the work, the better the training signal.

I am trying to find a precedent for this arrangement and I am struggling. The closest I can get is the long history of apprentices training their eventual replacements, but an apprentice was a human being who then took up the trade. The analogue here would be if the apprentice were a system that could be copied infinitely, worked continuously, and paid nothing. Which is, of course, what an AI agent is.

There is a word for a situation in which you are required, as a condition of keeping your job, to produce the means of your own redundancy, and in which the more effectively you do so the more quickly it arrives. I do not think it is a word anyone at Meta is using in the internal documents. I think the internal documents probably use words like "velocity" and "leverage" and "force multiplier." But the word that fits is coerced.

What to watch

A few things follow from taking this programme seriously.

First, the Meta programme will not remain at Meta. The economics are too attractive and the template is now public. Expect to see comparable programmes announced, or, more likely, not announced but leaked, at other large tech employers within twelve months, and then at large non-tech employers within twenty-four. The companies that move first will frame it as innovation. The companies that move last will frame it as competitive necessity. The substance will be the same.

Second, the US/EU divergence will widen, and it will become a deployment-strategy variable that firms discuss openly in earnings calls. This is already happening in content moderation, in data retention, in advertising. Workplace AI training will join the list. The likely result is that European workers end up, for reasons having nothing to do with technical merit, in a better position than American workers, their jobs made modestly less precarious by a legal architecture that their American counterparts lack. The American debate about whether such an architecture is worth building will, I suspect, sharpen.

Third, watch for the first serious organising response. Meta is not a unionised workplace, but tech-worker organising has grown through the 2020s, and this is precisely the kind of issue, intimate, clearly documented, easy to explain, that mobilises workers who were previously uninterested in collective action. The programme hands organisers a near-perfect framing: your employer is watching you in order to replace you. If tech organising gets its first major breakthrough at Meta, or at a Meta competitor following the same playbook, the Model Capability Initiative will be remembered as the inflection.

Fourth, and most important: the question of consent to training data, already contested in the creative industries, is about to become a live employment-law question in a way it has not been. "Can my employer use my work product to train a system that will replace me?" is a question courts and legislatures have not meaningfully answered, because it has not been posed this directly before. It is being posed now. The answers will shape the next decade of what AI deployment looks like at work.

I will say plainly where I stand on the underlying question. I do not think a worker being observed in granular detail, without the structural ability to refuse, in order to generate training data for their own replacement, is participating in something that deserves to be called consent. I think the arrangement is coercive, and I think describing it as coercive is not rhetorical, it is accurate. The fact that it is legal in California does not make it less coercive. It makes California's employment law less protective than it ought to be.

That is, in the end, what a programme like Meta's tells us. Not primarily about AI. About the shape of the floor beneath the people who work.

Footnotes

¶