Layoffs by Algorithm: Meta Faces Lawsuit Alleging Its AI Selected Employees on Medical and Maternity Leave

The Line That Made Me Reread the Complaint Twice

I’ve read a lot of AI-related mass-layoff stories on this blog. This is the first one that made me stop and reread a specific passage twice: one employee says his own manager warned him directly — taking the medical leave he was legally entitled to would “flag him in the system” as a layoff candidate. He took it anyway. He got laid off.

That’s one of the accounts in the lawsuit filed by 26 current and former Meta employees against the company and Mark Zuckerberg in federal court in California. The central claim: Meta’s internal AI system selected these workers for a round of roughly 8,000 layoffs specifically because they were on medical, maternity, parental, or disability leave — all legally protected categories.

I need to be upfront about one thing before going further: this is still an allegation, not a ruling. A federal judge denied the employees’ emergency injunction request to block the layoffs in July, and Meta has stated publicly that “workforce management and organizational decisions were and are made by people, not AI.” The case is proceeding to arbitration. But the technical details described in the complaint — true or not in the final outcome — expose an engineering problem I think is genuinely worth discussing, regardless of how this specific case resolves.

How the So-Called “Second Brain” Works

To understand the allegation, it helps to understand the architecture described in the complaint. Meta runs an internal assistant called Metamate, connected to agents dubbed “Second Brain,” trained to capture each employee’s routine:

Reading internal messages and documents.
Tracking keystrokes, screen activity, mouse use, and browsing history — captured directly from company-issued devices.
Dashboards measuring each employee’s AI token consumption.
Algorithm-assisted ranking and calibration tools that, according to the complaint, have already replaced much of the evaluation once done by human managers.

Based on this volume of surveillance data, the system generated a productivity score — and that score, according to the plaintiffs, is what determined who made the final layoff list.

The Problem I Find More Serious Than the Case Itself

What bothers me most about this story isn’t even the intent — I don’t believe anyone programmed the system to deliberately target people on leave. It’s the basic engineering flaw the complaint describes.

If a productivity score is calculated from keystroke volume, messages sent, and tokens consumed, anyone away for surgery, maternity leave, or temporary disability is going to generate, mathematically, a flat line of zero activity. Not because they’re producing less. Because they’re not there to produce, with legal approval to be absent.

And if the model doesn’t have a variable programmed to recognize “this person is out on approved medical leave,” it doesn’t see justified absence. It sees insufficient performance.

What the system measures	What the algorithm interpreted	The employee’s actual reality
Zero keystrokes	Inactive employee, no output	Temporary leave for medical treatment or surgery
Zero messages sent	Lack of engagement with the team	Maternity or paternity leave
Low productivity score	Priority candidate for the layoff list	Worker protected by federal leave law

This is what engineers call a poorly designed proxy metric: you use an easy-to-measure number (keystrokes, messages, tokens) as a stand-in for something hard to measure (someone’s actual value at work). It works fine until it runs into exactly the kind of exception the metric never accounted for — and in this case, that exception is protected by federal law.

Not the First Time

According to the complaint itself, this episode adds to an uncomfortable string of algorithmic surveillance controversies inside Meta: a click- and screen-tracking program justified as collecting training data for internal models, and a security failure that exposed confidential performance reviews to the entire company. It’s not an isolated incident — it’s a pattern the lawsuit is trying to document.

What I Actually Think

I don’t want to convict Meta based on an allegation that hasn’t been ruled on — that would be as careless as the very engineering mistake the lawsuit describes. But what stays with me is this: even if this specific case doesn’t hold up, the structural problem still exists in any company that decides to outsource layoff decisions to a number generated by digital surveillance.

A proxy metric isn’t the villain by itself — it’s a useful tool when well-designed and used as one input among others, not the only one. What turns it into a disaster is when the metric fully replaces human judgment, and nobody stops to ask “does this score actually capture what I want to measure, or just what’s easy to measure?”

When a company hands off the decision to lay someone off to a number that can’t tell laziness apart from approved maternity leave, the failure isn’t the technology’s. It’s whoever trusted it without questioning it.

I’m Left With This Question

Regardless of how this specific case is resolved in court, it leaves an open question for any company using AI in decisions about people: does your metric know how to tell justified absence apart from low performance — or does it just know how to count keystrokes?

Email: fodra@fodra.com.br
LinkedIn: linkedin.com/in/mauriciofodra

26 employees. An allegation not yet ruled on. An engineering flaw that, proven or not in this particular case, will keep existing in any system that confuses absence of data with absence of value.