The training data was the workforce

The leaked audio is short, and the line that has travelled is short too. Mark Zuckerberg, addressing Meta staff at an April all-hands, describing the company's Model Capability Initiative — the internal programme logging keystrokes, mouse clicks, and screen captures from employees using Gmail, VSCode, and Meta's internal collaboration tool Metamate — as a way to teach AI "how smart people use computers."¹ Published by the worker-advocacy outlet More Perfect Union on 22 May, the recording surfaced in the same news cycle as Meta's announcement of approximately 8,000 layoffs and the reassignment of 7,000 more employees into AI-focused roles.²

The temptation is to read this as a gaffe — a CEO caught saying the quiet part loud. I don't think that's what's actually going on here. What the audio captures is not a slip. It is an unusually candid description of a programme that was designed coherently, with its consent architecture, its jurisdictional carve-outs, and its substitution logic all working together. The scandal isn't that Zuckerberg said it. The scandal is that, said out loud, it describes exactly what the programme is.

The two events are one event. Meta is, by the company's own account, conducting the largest internal restructuring in its history. Roughly 10.8% of the workforce is being cut. Another 7,000 are being moved onto AI teams. And the productivity tools those workers, and the ones being let go, used every day were, at the same time, instrumented to capture the behavioural signal that would train the systems Meta is now building around. Zuckerberg's framing in the audio makes the relationship between these facts unusually explicit: the value of monitoring engineers is precisely that they are smart people doing knowledge work, and the AI needs to learn how that work gets done.¹

This is not "AI washing" in the usual sense — the practice of dressing ordinary automation in AI language to attract capital. It is closer to the opposite: rare candour about the substitution logic. Workers were monitored not to improve their own conditions or productivity, but to extract the signal that could eventually substitute for them. The data and the redundancies are two outputs of the same project. Calling that anything other than what it is would require ignoring what the CEO himself reportedly said about it.

I want to be careful here about what the evidence does and does not show. The Register described the audio as "purported," and Meta has not publicly confirmed its contents.¹ Reports also note Zuckerberg's claim that the data is stripped of identifiers, not used for performance management, and accessed only by AI systems rather than human reviewers.³ Those are meaningful caveats. They do not, however, touch the core fact, which is that Meta itself has not denied the programme exists, and the programme's design, what it collects, from whom, and to what end, is the thing under scrutiny. Whether the data is later anonymised is a downstream question. The upstream question is whether the workers being monitored knew, consented, or had any meaningful ability to refuse.

The carve-out tells you everything

EU-based employees are exempt from MCI data collection. This is the detail I keep returning to, because it is the cleanest piece of evidence in the entire story. Meta is a multinational. It runs Metamate, VSCode, and Gmail for staff in every jurisdiction it operates in. And it built, into the design of MCI, an explicit carve-out for workers covered by the General Data Protection Regulation.¹

That carve-out is not philosophical. It is regulatory. The European Data Protection Board has held, repeatedly, that covert keystroke logging fails the legitimate-interest test without explicit, freely given employee consent — and that consent in an employment relationship is presumed not to be freely given, because of the power asymmetry between employer and worker.⁴ Meta's lawyers will have read those guidelines. The company's behaviour tells you what they concluded: that MCI, as designed, would not survive contact with EU consent law. So EU workers were left out.

The standard defence here, and it is a reasonable one as far as it goes, is that operating differently in different legal jurisdictions is what multinationals do. GDPR is exceptionally strict; the US has nothing equivalent at federal level; running compliant programmes in each market is not, by itself, evidence of bad faith. I think that defence holds for, say, marketing data practices, where the consent threshold is genuinely contested. I don't think it holds here.

The reason it doesn't hold is that the carve-out demonstrates the company knew. The same programme, applied to the same kinds of workers, doing the same kinds of jobs, was deemed legally untenable in one jurisdiction and rolled out in the others. The non-EU workforce wasn't given a different version of MCI with stronger consent protections. They were given the version that wouldn't pass in Europe. The legal floor in the US and elsewhere is what permitted the design, and the design was optimised to that floor.

Where AI training data comes from workers, and where consent law is weak, the two facts tend to appear together. That is not coincidence.

This is the distributional pattern that keeps showing up in how AI data extraction operates. The workers whose behaviour gets captured are the ones with the least legal protection against capture. Kenyan content moderators training large language models. Filipino data labellers. American engineers whose keystrokes feed MCI. The European staff at the same company are spared not because their employer respects them more, but because their legislature insists on it. Take the legislature out of the picture and the respect evaporates.

The consent that wasn't asked for

I want to spend a moment on what consent would have looked like, because the absence of it is doing a lot of work in this story.

A meaningful consent regime for something like MCI would have required Meta to disclose, at the point of collection, that keystrokes, mouse movements, and screen captures from Gmail, VSCode, and Metamate were being logged for the explicit purpose of training AI systems that might in future be used to perform similar work. It would have required a real opt-out — not the kind that's technically available but professionally suicidal to exercise — and it would have required telling employees who else inside the company could access the data, even in derived form, and for how long it would be retained.

None of the reporting suggests any of this happened. Zuckerberg's own reported framing — that Meta deliberately did not explain the programme publicly so as not to tip off competitors³ — implies the inverse: that the absence of disclosure was a feature, not an oversight. The workers at the most monitored end of the programme appear to have learned about it from a leaked recording, which is to say, from the press.

That is the consent architecture. It is not an accident. It is the part of the design that did the most work to make the rest of the programme possible.

What UTAW is doing matters

The United Tech and Allied Workers union has begun organising around the disclosure. UTAW is a relatively new formation within the UK's Prospect union, recruiting white-collar workers in tech and professional services. The reporting describes the action around MCI as the first major white-collar AI labour response of 2026.²

I think this is worth taking seriously even though it is small and early. The dominant assumption about white-collar AI displacement — across the policy literature, across the consultancy reports, across the more optimistic boardroom slides — has been that knowledge workers will adapt, retrain, and reabsorb into higher-value roles, and that organised labour resistance of the sort that shaped industrial transitions of the twentieth century will not feature significantly. White-collar workers, the assumption goes, are not union workers. They will accept transitions individually rather than collectively.

UTAW is an early piece of evidence that this assumption may not hold. Whether the action gains traction is genuinely uncertain. The UK tech labour market is not large, white-collar union density is low, and Meta is not particularly exposed to UK-specific pressure. But what's significant is not whether UTAW wins this round. It's that the disclosure produced an organised response at all, framed explicitly around the surveillance-to-training-to-substitution chain. That framing is now in circulation. It will be available to the next group of workers who notice the same pattern at their employer.

No equivalent action has yet emerged in the US tech sector at comparable scale, and the structural reasons are familiar: weaker labour law, at-will employment, the cultural code of the US tech workplace. But the framing now exists in English, in a jurisdiction where labour organising is legally protected, attached to a named programme at a named company with a leaked CEO recording attached. That is the kind of artefact organising campaigns build around.

What the story is actually about

The question that keeps getting asked about AI deployment in knowledge work is whether it will create as many jobs as it destroys. I think this is the wrong first question, and the Meta case shows why. Before you get to the displacement, you have the data extraction that made the displacement possible. And the data extraction was conducted on the workers being displaced, without their meaningful consent, in jurisdictions selected for the weakness of their consent law.

That is not a side effect. It is not a regrettable feature of an otherwise neutral process. It is the design.

I read the leaked audio and I don't hear a CEO who has been caught doing something he is ashamed of. I hear a CEO describing a programme he believes is justified by competitive necessity, to an audience he assumed would understand the framing. The framing is: smart workers' behaviour is a valuable training input, and the way to get it is to log them while they work. The 8,000 leaving and the 7,000 reassigned are not collateral. They are the population from which the signal was drawn.

What there is to watch from here: whether the audio's authenticity gets formally contested, whether UTAW's action generates a regulatory response in the UK or EU, whether other large tech employers disclose comparable internal programmes, and whether US-based workers find a vehicle through which to push back. I would not predict any of these with confidence. I would say that the framing — that surveillance, training data, and substitution can be the same programme under the same name — is now in the room. It will not leave.

Footnotes

¶