Twenty-six people who used to work at Meta filed suit in Oakland this week, and the core of their complaint is not that they were laid off. It is how the list of names got made. They allege the company did not sit managers down to weigh the work in front of them. Instead, according to the 71-page complaint, Meta ran employees through a set of internal systems that scored, ranked, and selected them, and the scoring quietly punished anyone who had been away from their desk for a medical reason.
The complaint names the machinery. There is a system called Metamate, employee-trained agents described as "second brains," dashboards tracking AI token usage, keystroke and activity monitoring, and algorithmically assisted performance ranking and calibration. The plaintiffs filed anonymously, they come from six states, and they say Meta violated federal and state laws that protect workers who have disabilities, take medical leave, or are pregnant. Meta's answer, in full, is that "these claims lack merit and are not based on facts. Workforce management and organizational decisions were and are made by people, not AI."
Hold that sentence up to the light, because it is doing more work than it looks like. Nobody in this dispute claims a server sent out the termination emails on its own. The question the plaintiffs raise is narrower and harder: if a human being signs off on a ranked list that a scoring system produced, and the human does not know or cannot see why the scores fell where they did, who made the decision? Saying "a person clicked approve" is not the same as saying "a person judged the work." A rubber stamp is still a human hand, and it decides nothing.
The metric penalty
The most convincing part of the complaint does not require any villainy at all. It requires only that the system worked as designed. If you rank people on output, productivity, and volume of activity, then anyone who was absent for three months produces less measured output over the window, not because they are worse at the job, but because the window counted their empty weeks against them. The plaintiffs put it plainly: the scoring not only failed to account for protected leave, it penalized it. Someone who took twelve weeks for cancer treatment or a difficult pregnancy shows up in the data as a person with fewer commits, fewer tokens, fewer keystrokes, thinner numbers all around. The model does not hate them. It just cannot tell the difference between a person who did less and a person who was legally allowed to be gone.
This is the generalizable problem, and it is why the case matters beyond one company. Any output metric, fed to any ranking system, will do this unless someone deliberately corrects for protected absence. That correction does not happen by accident. It happens because a person who understands the law builds it in, tests for it, and audits the result. The complaint alleges Meta did none of that, and it points at recently adopted California and New York City rules that require exactly that kind of bias testing on automated employment tools.
The surveillance that fed the scores
Sitting underneath the ranking is a second story that deserves its own attention. Earlier in the year, the plaintiffs say, Meta stood up a monitoring program that captured keystrokes, screen content, mouse movement, browser history, messages, emails, and voice, video, and location data from company devices, and used that data to build AI tools. According to the complaint, the program was announced through a low-visibility internal post written by an engineer, not a senior leader, dropped in a secondary group rather than the official employee-notice channel. On some teams there was no consent prompt at all, and at first no way to opt out.
That is the pipeline worth watching. Surveillance data becomes training data becomes the scores that decide who stays. Each step is defended as ordinary on its own. Companies monitor devices. Companies build tools. Companies rank performance. Stacked together, they describe a workplace where the record of your every action is quietly assembled into a number that can end your job, and you were told about it in a post you probably never saw.
Why this is in court at all
Meta's employment agreements push workplace disputes into individual private arbitration, which normally keeps stories like this out of public filings entirely. The plaintiffs found the seam. Those agreements, they argue, do not cover requests for temporary relief, so they went to federal court asking Judge William Orrick to block Meta from finishing the layoffs, which were set to begin on July 22, while the underlying claims proceed in arbitration. It is a procedural move, but it is the reason any of this is readable at all rather than sealed behind a confidentiality clause.
None of it is proven. A complaint is a set of allegations written by one side, and Meta's denial is a claim too, not a finding. What tilts my read is that the plaintiffs are not asking anyone to believe in a rogue algorithm. They are describing a system behaving exactly as an output-ranking system behaves, and a defense that leans on the word "people" without explaining what those people actually saw. The closest precedent, the Mobley case against Workday, already cleared the way for a collective action over alleged AI hiring bias, so courts are no longer treating "a human was technically involved" as a full answer.
The document that will settle this is not either side's statement. It is whatever exists inside Meta showing what the scoring actually weighed, whether the leave windows were flagged or ignored, and what the approving managers were shown before they signed. If that record shows the metric window swallowed protected leave without correction, the phrase "made by people, not AI" will read less like a defense and more like a description of who was standing nearby when the machine did the sorting.