Meta was recording its own employees' screens and key presses to train AI, and that data spilled to the whole company. The program logged mouse moves, clicks, keystrokes, and whatever was on a worker's screen. Then the databases holding all of it were left open to anyone inside Meta. On June 22, 2026, Meta paused the program to investigate.
Meta started rolling out the program, known internally as MCI, to US staff in April 2026. For most of them it was not optional. The point was to teach Meta's AI to use software the way real people do, with real workers as the examples.
What spilled out
The leak went past clicks and key presses. According to Business Insider, the exposed data also held private chats, performance data, and transcriptions. A Meta engineer flagged it as SEV 2, on a scale where 0 is the worst. So it was serious, but not the top rank.
Meta says no harm is confirmed yet. A spokesperson said the program was "carefully designed" with "privacy safeguards," and that there is "no indication" any worker's data was improperly accessed. Meta says it found the problem on June 18 and pushed a fix within four hours. But the fix did not hold. The data had to be locked down again, and one former employee called it "a mess."
Business Insider on the leaked data, WIRED on Meta's response
Why workers were angry
Workers saw it coming. At launch, no one could opt out. After protests, Meta let them pause the recording for up to 30 minutes and ask to be dropped from the program. More than 1,600 staff had signed a petition against it, and UK workers started a union drive over the program.
Recording your own staff to train AI was always going to cost trust, and now that the same program has exposed their private data across the company, that trust looks gone. Meta executive Stephane Kasriel told staff that MCI will not turn back on until the company trusts its data controls. He also said Meta had already "gathered sufficient data to assess the long-term value of the tool." So even Meta sounds unsure.
