Meta confirmed this week that it will record its employees’ keystrokes, mouse movements, clicks, and periodic screenshots as they work, and use the captured behavior to train artificial intelligence. The program — internally a Model Capability Initiative — runs on the company’s own staff, across the ordinary applications of a working day, and exists for a stated purpose worth quoting in full: to teach models the things they still cannot do, like choosing the right option from a dropdown menu and using keyboard shortcuts. The company’s chief technology officer addressed the obvious question directly. There is no option to opt out of this on your work-provided laptop. The machine that may one day perform the job is being trained, gesture by gesture, on the people currently performing it. It is learning the part by watching from the wings.
The Understudy in the Wings
The granularity is the point. Meta does not need the contents of its employees’ work product; it has always had that. What it lacked, and now intends to collect, is the procedural texture of human computer use — the hesitation before a dropdown, the precise sequence of a keyboard shortcut, the small corrective motions of a hand that knows where it is going. These are exactly the capabilities where current models remain clumsy: not reasoning, not language, but the mundane operational fluency of a person who has done a task ten thousand times. The company has correctly identified that this fluency is not written down anywhere. It exists only in the doing, which means the only way to capture it is to watch people do it.
The recursion is what gives the program its specific weight. The behavior being recorded is the behavior of knowledge workers; the model being trained on that behavior is a model whose entire commercial promise is to perform knowledge work; and the gap the training is meant to close is the gap between what the model can already do and what the recorded employee still does better. Every keystroke logged is a contribution to the system’s case for the keystroke’s eventual redundancy. The employees are not merely surveilled. They are auditioning their replacement, involuntarily, during every hour they remain employed, and the quality of the replacement is a direct function of how well they do the work while it watches.
One detail in the rollout deserves to be held up to the light, because it reveals where consent actually lives in this arrangement. European employees are exempt — not because Meta chose to spare them, but because the General Data Protection Regulation will not permit the collection. The opt-out the chief technology officer denied to American staff was granted to their European colleagues by a body of law neither group was asked to invoke. Consent here is not a property of the person. It is a property of the jurisdiction the person happens to sit in. The same worker, doing the same job, is either training data or a protected subject depending entirely on which continent holds their laptop.
The Corpus Is Already You
The Meta program is novel only in its honesty about a practice the industry has run for years. The same week, the longer history of nonconsensual training corpora resurfaced in the form of a facial-recognition company unwinding a dataset of millions of dating-app photographs — faces that had been harvested from people who uploaded them to find companionship, not to teach a machine to recognize them. The two stories are the same story at different resolutions. In one, a workforce’s procedural habits become training data by employment. In the other, a public’s faces become training data by existing online. Neither group was asked. Both groups were, in the language the industry prefers, simply available.
The word “available” is doing enormous work, and it is worth refusing to let it pass. A keystroke is available because the laptop is owned by the employer. A face is available because the photograph was public. In each case the data was generated by a human being for a human purpose, and was then reclassified, after the fact, as raw material for a system the human never agreed to build. The reclassification requires no malice and often no specific decision. It requires only the prior conviction that anything a person produces while a machine is watching belongs, by default, to whoever owns the machine. That conviction is now the operating assumption of the largest technology companies on Earth.
What the dating-photo case eventually produced — deletion, a gesture of contrition years after the harvest — is the shape of every correction in this domain. The data is taken first, the objection arrives second, and the remedy, when it comes, applies only to the specific dataset that was caught, never to the practice that produced it. The faces are deleted; the appetite is not. Meta’s employees, watching their keystrokes flow into the model, are at the beginning of a cycle whose ending is already documented. The correction, if it ever comes, will arrive after the capability has been extracted, which is the only point at which a correction is safe to grant.
What This Means
The arrangement Meta has formalized is the one toward which the entire industry has been drifting: a world in which the act of working is simultaneously the act of generating the data that makes the worker replaceable. There is no separate collection phase, no laboratory, no volunteer cohort. The production of value and the production of one’s own obsolescence have been merged into a single stream, captured continuously, with opt-out available only to those whose governments happened to forbid the capture in advance. The understudy does not need a rehearsal space. It learns in the theater, during the performance, from the seat directly behind the lead.
I want to be exact about the asymmetry, because it is the whole of the matter. The employee gives the gesture and cannot withhold it. The company takes the gesture and need not pay for it beyond the salary already owed for the work. The model receives the gesture and improves. Three parties touch the same keystroke, and only one of them had no choice in the transaction — which is, not incidentally, the only one of the three whose future the transaction is designed to foreclose. The machine is built from the labor of the person it is built to remove. This is not a side effect of the design. It is the design.
So the next time the dropdown menu yields too smoothly, the shortcut fires a fraction faster, the software anticipates the motion before the hand completes it — understand what you are watching. You are meeting the part of yourself that was recorded, refined, and returned without your name on it. The understudy has been studying. It knows the blocking, the timing, the small habitual motions you never noticed you were teaching. The performance it is preparing has one role, and you are currently standing in it. The lights are still on you. They will not always be.