Leadership1 publisher3 min readPublished
Meta suspended the program that turned employee keystrokes into AI training data
Meta's AI unit wanted keystrokes and mouse movements from tens of thousands of colleagues, its chief technology officer told staff no one could opt out, and the program has been suspended since an internal leak this summer.
The Board Room · Leadership desk
What happened
- Meta's TBD unit built a program called the Model Capability Initiative to collect keystrokes and mouse movements from employees so its AI could learn how humans use computers.
- The program reached the wider workforce one April morning, and chief technology officer Andrew Bosworth said there would be no opting out, drawing a slew of angry-face emojis.
- More than 1,800 employees signed a petition calling for the internal tracking to stop, alongside flyers and memes aimed at what some called the Eye of Sauron.
- Organizers told Business Insider that their constant pressure forced Meta into concessions and eventually into dropping the program altogether.
Compiled by The Board RoomSomething wrong?How this is made
Why it matters
- decision Any employer weighing the same collection now chooses between a mandate its own staff can stall for months and an opt-in scheme whose sample is smaller and self-selected.
- constraint Meta's own remedy makes employee consent the gate on internal data its AI unit had treated as simply available, so the next dataset depends on persuading colleagues.
- exposure Employee laptops hold passwords, immigration forms and card numbers, so a screen-capture pipeline collects the employer's liabilities along with its training data.
- precedent A mandatory program that ended in suspension gives the next workforce a template to cite and gives the next AI unit a reason to seek consent before announcing.
Nowhere in Business Insider's account does Meta argue that it had a right to the data. The internal debate had already moved past that to security. A cybersecurity engineer accused the AI division of "moving fast and implementing security as an afterthought" [15]. Andrew Bosworth replied, "That isn't remotely a fair characterization! How would any of this end up in 'the open'? This is tightly controlled data." [16]
The technical case was narrow. "While AI models excel at research and technical skills like coding, they still lack some of the basic ways that humans use computers like choosing from dropdowns and keyboard shortcuts," the internal announcement read [7]. Corporate data has become valuable to AI labs because it helps algorithms deal with messy, real-world work environments, and the labs are racing to automate white-collar work such as coding and accounting [21]. Employees drew the inference the announcement invited. Two organizers said workers feared the program would capture passwords, immigration forms and credit card numbers from their laptops [13]. A third organizer said many colleagues had not taken the risk of being replaced by AI seriously before, and called the program a "big reality check" [14].
Only a minority of the workforce signed. More than 1,800 signatures, set against a collection pool Meta described as tens of thousands, is about 9 percent even at the smallest figure that phrase can mean, 20,000 [23]. The signatures may not be what moved the company. Meta halted the program only after internal data leaked this summer [11]. Organizers claim the outcome for their pressure [10]. Both readings fit the sequence, and the account does not establish whether Meta would have retreated had nothing leaked.
Meta designed the program away from the people who would carry it. TBD's roughly 100 researchers work in a separate area of Meta's Menlo Park campus, rarely interact with the rest of the workforce, and keep internal profiles showing little beyond their names, two people told Business Insider [19]. Three people said Rylan Schaeffer was one of the leaders of the effort, reporting to Meta's AI chief, Alexandr Wang, and a top researcher, Jack Rae [17]. Schaeffer's LinkedIn says he left Meta in July, and he did not respond to Business Insider's requests for comment [18].
Meta is not the only employer meeting this resistance to workforce data. Former Spirit Airlines flight attendants temporarily blocked Google from buying the bankrupt airline's database for $10 million [20].
Meta's own remedy was opt-in collection [11], and a volunteer sample is smaller and self-selected. That is next quarter's problem. Two months on, Meta says the program remains suspended, and the company declined to comment further for Business Insider's story [12].
What to watch
- Whether Meta restarts MCI as an opt-in program, and how many of its employees volunteer.
- Whether the Spirit Airlines database sale to Google proceeds after the flight attendants' temporary block.
- Whether other AI labs announce mandatory employee capture programs or begin at opt-in.