Earlier this month Meta rolled out a software package called the Model Capability Initiative across employee machines. It captures mouse movements, clicks, keystrokes, menu navigation and periodic screen recordings, all of it framed internally as training data for Meta's next generation of AI agents, the systems the company is positioning as automated colleagues capable of completing the kind of computer-based knowledge work that currently fills the screens of the people being recorded. The rollout coincided, by a margin of weeks, with Meta sending layoff notices to roughly 8,000 staff, around 10 percent of the company. Another 7,000 are being reassigned into AI initiatives. The optics, several employees said, told themselves.
What has happened next is unusual for Meta. Flyers began appearing in US offices reading "Don't want to work at the Employee Data Extraction Factory?" and citing the National Labor Relations Act, the 1935 statute that gives most US private-sector workers the right to organise. A petition is circulating. One engineer quoted in coverage of the protests said simply, "I don't want to live in a world where humans are exploited for their training data." The framing matters: workers are not arguing that the software is buggy or invasive in the usual creepy-IT sense, they are arguing that their labour is being expropriated as a raw material for the very systems being built to replace them, and that this is a workplace rights issue rather than a privacy complaint.
The organising effort is more advanced in the UK. United Tech and Allied Workers, a branch of the Communication Workers Union, has confirmed a union drive at Meta's London offices, the first significant attempt to unionise the company anywhere in the world. UTAW's Eleanor Payne told reporters that "the workplace surveillance and training AI models is the No. 1 thing" driving employee interest. The UK's recently strengthened union recognition rules, passed last year, materially lower the threshold for forcing the company to recognise a bargaining unit. A separate UTAW statement framed the moment in language that captures what makes this story new: Meta's workers, the union said, are "facing devastating job cuts, draconian surveillance, and the cruel reality of being forced to train the inefficient systems being positioned to replace them."
That phrase, "forced to train the inefficient systems," is doing work. It accepts that today's AI agents cannot yet do most of what Meta's white-collar staff do, and identifies the gap as a labour issue rather than a technical one. The training data the company needs to close that gap is precisely the data its current employees generate. Refusing to provide it, individually or collectively, is one of the few forms of leverage white-collar tech workers have had in a year defined by mass layoffs across Amazon, Microsoft, Oracle, Cisco, Block, Cloudflare and now Intuit (covered separately in today's issue). Most of those cuts have happened without visible resistance. Meta is, so far, the exception.
Two things will determine whether this matters beyond Meta. The first is whether the UK union drive succeeds in winning recognition, which would establish a precedent for collective bargaining over the use of employee data in AI training across the European tech sector. The second is whether US labour law, designed for an industrial economy, can accommodate the argument that mandatory mouse-tracking for AI training is a "term or condition of employment" subject to collective negotiation under the NLRA. Both questions will move slowly. The layoffs and the data collection will not. Meta's 8,000 outgoing staff have already made their unwilling contribution to the training set, and the workers staying will continue making theirs each time they open a laptop. The protest is, in one sense, an argument that this exchange should at minimum be negotiated, not assumed. Whether the company will treat that as anything more than an inconvenience is, for now, an open question.