On Monday, 26 current and former Meta employees filed a 71-page complaint in the Northern District of California. They were all cut in the May reduction in force — the one that took roughly 10% of the company. They come from California, Florida, Illinois, New York, Pennsylvania and Washington. They do different jobs. They share exactly one trait, and it is the entire lawsuit: every one of them took, requested, or was approved for protected leave in the last 24 months.
Meta’s response, from a spokesperson: “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. We’re going to need it later.
The list
Here is the allegation, in the plaintiffs’ own words, and it is worth reading slowly:
“Meta did not assemble the termination list through the considered judgment of managers who knew the work. Instead, Meta used a constellation of internal artificial-intelligence systems — including a system referred to internally as ‘Metamate,’ employee-trained ‘second-brain’ agents, keystroke- and activity-monitoring data, AI-token-usage dashboards, and algorithmically assisted performance ranking and calibration — to score, rank and select employees for inclusion on the list.”
Read the inputs again. Keystroke data. Activity monitoring. AI-token-usage dashboards.
Somewhere inside Meta there is a dashboard that counts how many tokens you spent talking to the AI, and that number was allegedly load-bearing in whether you kept your job. Not what you shipped. Not what you decided. How much you used the product. If you have ever wondered what “AI-first” means operationally, this is a candidate answer: it means your consumption of AI became a proxy for your value, and the proxy got a column in the spreadsheet that fired you.
The mechanism the complaint describes doesn’t require anyone at Meta to have harboured a single unkind thought about pregnant engineers. It’s simpler and worse than that. The system scored people on accumulated output. You cannot accumulate output while you are on medical leave. That is what leave is. Nobody neutralised the inputs, nobody excluded the people on leave from the ranking, and so the ranking did the only thing it could do: it read absence as underperformance and handed back a list.
The individual stories in the filing are almost clinical. A scientist selected while on pre-birth pregnancy leave. A manager demoted after one medical leave, then picked for the list weeks into his second. An engineer whose rating was marked down for the “broken time” when an injury kept him off the keyboard. No villain required. The machine was working exactly as specified. The specification was the problem.
The part where the story eats its own tail
Regular readers may recall that in April, Meta was installing keystroke trackers on its employees’ laptops — mouse movements, clicks, screen snapshots — to train workplace AI agents. At the time the joke wrote itself: the workers were training their own replacements.
The complaint alleges the joke was insufficiently ambitious. That monitoring program captured keystrokes, screen content, mouse activity, browser history, messages, emails, and voice, video and location data on company devices. The data built the AI tools. The AI tools built the list. The list included the people whose keystrokes trained it.
The workers didn’t just train their replacement. They trained the thing that chose them.
On the consent question, the plaintiffs are blunt: the monitoring program was “announced through a low-visibility internal post — made by an engineer rather than a senior leader, in a secondary group rather than Meta’s official employee-notice channel — with little notice and no consent or click-through acknowledgment.” On some teams, no prompt at all. Initially, no way to opt out. If you’re going to build the instrument of your own workforce’s assessment out of their unwitting mouse movements, the least you could do is put it in the announcements channel.
”People, not AI”
Back to Meta’s statement. “Workforce management and organizational decisions were and are made by people, not AI.”
This is not a denial that the systems exist. It’s a claim about where the decision lives. And it’s the exact question the next few years of employment law will be spent on, because there is a difference between a human making a decision and a human signing one. If Metamate ranks 80,000 people, and a director approves the bottom decile without reopening the ranking, who decided? Legally, the human. Functionally, the model. The complaint’s requested remedy targets this precisely: they want an independent audit of the algorithmically assisted selection process, and their terminations halted until that happens. They want to open the box.
They also have a scheduling problem. Meta requires employees to sign mutual arbitration agreements with class action waivers, so these 26 must go individually — no class, no aggregation, 26 separate proceedings. Hence the request for a preliminary injunction to restore their employment status as of May 20 while it plays out. The case is before U.S. District Judge William Orrick.
The claims run through the FMLA, the ADA, the Pregnancy Discrimination Act, state leave statutes. Notice what isn’t there: no statute against being ranked by a machine. There isn’t one to invoke. The plaintiffs are reaching for civil rights law from the 1960s and 1990s because it’s the only lever in the room, arguing that an algorithm produced a disparate impact on protected classes. It’s the same theory now grinding through the Workday applicant-screening case, where a judge has already refused to dismiss. Discrimination law is being asked to do a job it wasn’t built for, because nothing was built for this.
What this means if you work anywhere
Don’t file this under “Meta problem.” File it under: your performance data is now a training set, and the metrics that feel most objective are the ones most likely to be doing something stupid.
Token usage is a perfect example of the failure mode. It looks like a real number. It’s countable, comparable, dashboardable. It measures nothing about whether you are good at your job. But it’s available, and in algorithmic management, available beats meaningful every single time. The metrics that get used are the ones that are easy to collect, and the ones easy to collect are the ones a keylogger can see.
Which points at the practical read for anyone whose work involves periods of not visibly producing — leave, deep thinking, mentoring, the meeting where you talked someone out of a bad architecture. None of it registers on an activity monitor. If a scoring system is in the building, and increasingly it is, your defence isn’t to type more. It’s to make sure the part of your job that matters leaves a trace something can count — and to know that a gap in the data will be read as a gap in you, by a system that has never once considered why you were away.
Twenty-six people are now asking a federal judge to make Meta show its work. The rest of us will find out what’s in the box when they do.
