Meta sued: when an AI scoring system for a layoff plan meets labor law

In plain terms

Twenty-six former Meta employees are suing the company. They claim that AI-assisted systems were used to rate and rank employees on a layoff list, and that these systems disproportionately disadvantaged people with disabilities, on medical leave, or pregnant. Meta disputes this. This is one of the first lawsuits that directly targets the tool rather than the decision.

Ce qui est allégué

The plaintiffs, who have filed the lawsuit anonymously, invoke federal and local laws prohibiting discrimination and retaliation against employees with disabilities, on medical leave, or pregnant. They also accuse Meta of not testing its AI systems for bias, in violation of recently adopted texts in California and New York.

According to the complaint, several internal systems would have contributed to rating and ranking employees: "Metamate," an assistant based on a large language model; a "second brain" trained on employee communications and documents; and a productivity score derived from keyboard strokes, screen content, emails, and browsing history. According to the complaint, the company would have relied on factors such as productivity and AI token usage to eliminate thousands of positions this year.

Meta responds unambiguously: "Workforce management and organizational decisions were and are made by people, not AI."

Under the hood

Meta's defense—that the decision is made by humans—is legally sound, and that's precisely what makes the dispute interesting. Because the question the court will ask is not "who clicked?", but "on what did the person who clicked base their ranking?" A human who validates a ranking produced by a biased scoring system does not neutralize the bias: they ratify it. Anti-discrimination law has long known this figure under the name of disparate impact—intent doesn't matter, only the effect measured on a protected group.

This is where the choice of variables becomes explosive. A productivity score based on keyboard strokes and time spent on screen is, almost by construction, a proxy for presence—therefore for absence for medical reasons. As for the use of AI tokens as a performance indicator, it is exactly the metric being discussed this week regarding engineer quotas: an easy figure to extract, correlated with a thousand things, causal of nothing.

So what

The tipping point is regulatory. The California and New York laws imposing a bias audit of automated decision-making tools transform an ethical debate into a documentary obligation: either the company produces the test, or it does not. This case will determine whether the obligation has teeth.

For any executive deploying HR scoring assisted by a model, the operational conclusion does not await the verdict: the trace of the bias audit is not a compliance formality, it's your only line of defense. Without it, you don't have a decision-making tool. You have a piece of evidence.