26 employees are suing Meta, alleging its internal AI systems (activity monitoring, token-usage dashboards, and performance rankings) skewed layoffs against those on medical or parental leave; it's being called the first major workplace algorithmic-discrimination case. Meta denies it: "people decided, not the AI."
For years the debate over AI's risks revolved around products: chatbots that hallucinate, fake images, buggy code. This week the frontline shifted to far more personal ground. CNBC and other outlets reported that 26 current and former Meta employees sued the company, alleging that its internal artificial-intelligence systems helped decide who to lay off in a cut of roughly 8,000 jobs —about 10% of the workforce, announced in May— and that they did so in a way that discriminated against workers who were on medical or parental leave. This isn't a lawsuit about what AI produces. It's a lawsuit about what AI decides about people.
When productivity scoring feeds the layoff list
According to the lawsuit, Meta leaned on a battery of automated signals to rank and compare its employees' performance: keystroke and activity monitoring data, AI token-usage dashboards, and algorithm-assisted performance rankings. The problem, the plaintiffs say, is that those metrics structurally punish anyone who had been absent for legally protected reasons: someone who took maternity leave, had surgery, or had a medical condition generates less measurable activity and, by design —according to the lawsuit's own wording— those scores "cannot be accrued" by an employee on protected leave. When that score feeds the selection of cuts, a leave shielded by law translates, in practice, into a higher probability of losing your job. Of the 26 plaintiffs, about half had taken caregiving or pregnancy leave: eight women on maternity leave, four men on parental leave, and one woman who cared for a family member and then took bereavement leave.
Four federal laws in the crosshairs, and a one-sentence defense
The lawsuit invokes heavy artillery from US labor law: the FMLA (which protects medical and family leave), the ADA (which protects people with disabilities), the Pregnancy Discrimination Act, and the Pregnant Workers Fairness Act. The argument is that an automated system isn't exempt from those laws just because it's an algorithm: if the outcome discriminates, it discriminates, regardless of whether a manager or a model signed off on it. Meta, for its part, responded with a compact and revealing defense: it maintains that the decisions "were made by people, not AI," and that the claims "lack merit and are not based on facts." That sentence is precisely the battlefield of the case, because the question that will decide the suit isn't whether there was a human at the end, but how much that human really weighed against a list the machine had already ranked.
Why this case opens a door that won't close
Although the use of AI in human resources has been growing for years —résumé filtering, performance evaluations, attrition prediction— nearly all prior litigation was theoretical or low-profile. This one is different for three reasons: the size of the defendant, the explicitness of the accusation (it isn't a diffuse bias, it's "the score penalized a protected leave"), and the fact that it targets AI applied to decisions about people, not about products. Win or lose, it sets the terrain: from now on, any company that uses automated scoring to decide hires, promotions, or layoffs knows the judge's question will be "show me the trail." Who scored whom, with what data, who reviewed the list, and what changed after that review. Algorithmic discrimination has stopped being an academic panel and become a lawsuit with a name attached.
What it means for anyone building with AI
The lesson isn't "AI for decisions is bad." It's more uncomfortable and more useful: the moment an automated system touches a decision that affects a person, the value stops being solely in the outcome and shifts to being able to demonstrate how you got there. A score with no trail is indefensible; a score with a human who genuinely reviews it, with a record of what they saw and what they decided, is another thing entirely. At NeuralOS we start from that same doctrine for automations and agents: sensitive actions go through explicit human approval before they run, and every step is recorded in an audit-log —who, what, when, with what data— instead of disappearing inside a black box. We don't say this as a slogan, nor do we promise it will shield you from a lawsuit; we say it because Meta's case teaches, with no smoke, that the "human in the loop" only counts if you can prove they were there. Automating is easy. Being able to explain the decision —that's the real work.