Meta wanted to train its AI models by recording how employees used company software. Two months later, it announced a pause after sensitive material—including private conversations and personnel information—became accessible to staff across the company.

The program began in April, when Meta started installing software on US employees' computers to capture mouse movements, clicks, keystrokes, and occasional screenshots in selected work applications. A spokesperson said the data would train models to operate computers, rather than assess staff performance, and that safeguards would protect sensitive content.

Staff objections prompted Meta to add temporary pauses and exemption requests in June. Later that month came the announced security pause.

Why the Access Was a Problem

A recording creates another copy of information that might already be protected in its original system. A personnel record can have carefully restricted permissions while a screen capture of that record is stored somewhere with broader access. The team collecting training examples has to protect the captured information too.

That is why need-to-know access restrictions matter even when no misuse has been identified. The training purpose does not give everyone working on AI a reason to read private conversations or personnel information. Access should follow the actual task, with unrelated sensitive content excluded or removed before the recordings are made available for training.

Who Controls the Expertise and Who Benefits?

Along with sensitive information, Meta was collecting demonstrations of how work gets done. An ordinary task had two outputs: the finished work and an example that management could reuse to develop automation.

The person supplying that example did not necessarily control its future use. A temporary pause or exemption request offered some influence over collection, but left wider questions unanswered: who could authorize reuse, whether staff could decline participation, and how they would be involved in decisions about automating their jobs.

Those decisions determine who benefits. A tool that removes tedious steps could give people more time for difficult cases. Management could also use automation to redesign roles or reduce staffing. The reporting does not establish that Meta used this program to replace particular workers, but people have a stake in those choices before the outcome is known.

Meaningful participation would require management to explain the intended uses and give staff a role in discussing them. Notice that a recorder is running does not answer what will happen to the expertise it captures.

Does Recording the Work Teach AI to Do It?

Even with agreement about collection and use, a technical question remains: what can a model learn from the visible actions?

Consider a hypothetical support specialist who stops an account change after noticing that a request conflicts with an earlier conversation. A recording could show the specialist opening a record, reading a message, and declining to proceed. It might not explain which conflict mattered, what authority the specialist had, or what evidence would have justified approval.

Can the model handle a new request, distinguish an exception from the usual procedure, and seek help when the evidence is insufficient? Replaying the demonstrated sequence would not establish those abilities. The available reporting does not show how successfully Meta's models learned from these recordings.

Experienced staff can help design those tests and recognize plausible actions that would be wrong in context. Involving them is therefore part of evaluating the system, as well as a question of having a say in how their work is used.

The most valuable thing a specialist did might have been recognizing when not to click. Someone training an AI to do that job needs to ask why.