The training set is you

Meta has begun logging the mouse movements, keystrokes, and periodic screen captures of its US employees. The programme is called the Model Capability Initiative. According to reporting by Tech Policy Press and CNBC late last week, the software is installed on every US work device, runs continuously through the working day, and feeds the captured data into the training pipeline for Meta's internal AI agents.¹ European employees are excluded because European law would not permit it. Meta has told staff that the data will not be used to inform performance management.

I want to take that last sentence seriously for a moment, because it is doing a great deal of work.

What is actually being collected. Keystrokes are not a summary of work. They are work, captured at the resolution of intent: the drafts you wrote and deleted, the Slack message you started composing to a colleague and thought better of, the password you mistyped, the search query you ran before you knew what you were looking for. Periodic screenshots capture whatever is on screen, including the windows you have open that are not strictly job-related, the personal email you checked at lunch, the medical portal, the union forum. Mouse movement telemetry, in combination, reconstructs hesitation, distraction, the rhythm of a person thinking.

This is the most granular workplace surveillance regime ever assembled at scale by a major US employer, and it is being installed not to manage workers but to replace the cognitive texture of their jobs with a model.

Who is being trained on whom. Meta is not the first company to record what employees do at their desks. Call centres have done variants of this for two decades. Amazon warehouses time bathroom breaks. What is new is the purpose. The older surveillance regimes existed to discipline the worker who was being watched. The Model Capability Initiative exists to extract the worker's judgement, encode it into a model, and then deploy that model into the same workflow. The worker is not the subject of the surveillance. The worker is the training set.

This is a meaningful shift, and I do not think it has been named clearly enough in the coverage so far. When your keystrokes are used to manage you, you remain the locus of value, however coercively. When your keystrokes are used to train the system that will sit alongside you and eventually in your seat, you are being asked to participate in the construction of your own redundancy. The consent question, such as it is, has been collapsed into the employment contract.

Why Europe is excluded. The GDPR's purpose-limitation principle, its rules on processing employee data, and the works-council consultation requirements in countries like Germany and France would make this programme difficult to run lawfully. Meta knows this; that is why European employees are out of scope. The same company, the same software, the same business case, the same workers doing the same jobs, and a different legal regime produces a different outcome.

That number is the policy argument. It is not that surveillance of this depth is technically unavoidable, or that AI training inherently requires it. It is that one jurisdiction has decided workers have a say in what gets done to them at work, and another has not. The asymmetry is the evidence.

The consent that isn't. Meta's US employees were not asked. They were informed. The distinction matters because the standard defence of this kind of programme is that employment is voluntary and the terms are disclosed. But voluntary in what sense? The labour market for senior engineers is tight, but it is not so tight that walking away from Meta over a monitoring policy is costless, and it is much less tight further down the org chart. The contractors, the support staff, the people on H-1Bs whose immigration status is tied to the job, their consent is the most theoretical of all.

I find the framing that this is about productivity, or about AI capability, less interesting than the framing that it is about who gets to refuse. European workers can refuse, through their works councils and their data protection authorities and the consultation rights baked into their employment law. US workers, at Meta and increasingly elsewhere, cannot. The same software arrives at both desks. Only one desk has a no.

What to watch. Whether this programme stays inside Meta or migrates to its peers; whether the "not for performance management" commitment survives the first quarter in which a manager wants to use it; whether any US state legislature responds; whether the workers themselves organise around it. The technology is not the variable here. The technology is settled. What is unsettled is whether the people being recorded have any standing to say no.

Footnotes

¶