The plaintiffs say AI tools scored their productivity while they were on protected leave, then used that score to decide who got cut
A federal judge has declined to stop Meta from proceeding with layoffs of 26 employees who allege the company used AI tools to select them for termination in a way that discriminated against workers on protected medical, parental and disability leave. The ruling is far from a clean win for Meta, though.
US District Judge William Orrick in Oakland, California, denied the plaintiffs' request for a temporary restraining order (TRO) that would have paused their separations, which begin July 22, while the underlying claims are arbitrated. At the same time, Orrick's written order acknowledged the case raises "serious questions going to the merits," and he left the door open to revisiting his decision once more evidence comes in.
How AI tools fed into the layoff selection
The dispute traces back to Meta's April 2026 announcement that it would cut roughly 8,000 jobs, about 10% of its global workforce, as part of a push to "remake itself into an 'A.I.-first' company," according to the plaintiffs' complaint. The 26 plaintiffs, proceeding anonymously as Does 1 through 26, filed suit on July 13, alleging Meta relied on a "constellation" of internal AI-assisted systems to score, rank and select employees for its reduction in force (RIF).
Those systems reportedly included Metamate, an internal large-language-model assistant; "second brain" agents trained on employees' own communications and documents; productivity scoring drawn from keystrokes, screen content, mouse activity, browser history, messaging and email data pulled continuously from company-issued devices; internal AI-token consumption dashboards; and AI-assisted performance review tools, per the complaint.
As HRD has previously reported on the underlying lawsuit, plaintiffs claim these systems weren't paused for employees on approved leave. As a result, their AI-adoption and productivity scores allegedly dropped while they were legitimately out of the office, feeding directly into the termination selection. The complaint brings 21 causes of action spanning the Family and Medical Leave Act, the Americans with Disabilities Act, Title VII, the Pregnancy Discrimination Act, the Pregnant Workers Fairness Act, and a string of state-level leave and disability statutes in California, Washington, New York and elsewhere.
Meta disputes nearly all of it. In opposing the TRO, the company submitted a declaration from Linh Doan, its director of HR business partner enablement, who stated that selection decisions were made by human business leaders using documented, neutral criteria: job level, historical performance ratings, tenure, location, job function, specialized skills, and reporting-structure efficiency. Doan's declaration states that no plaintiff was selected based on leave status, disability, or any other protected characteristic, and that no selection decision was made by AI.
Meta declined to comment on the ruling itself, though the company has consistently maintained, as Reuters reported, that the layoff decisions were made by people rather than by AI.
Weighing the standard four-factor test for injunctive relief, Orrick found plaintiffs had cleared the lower bar of showing serious unresolved questions on the merits but not a likelihood of success, given the sharply conflicting declarations on each side and the fact that discovery hasn't yet tested Meta's account. On irreparable harm, the order draws a clear line: lost health coverage, lost employment, the leave time itself, and unvested equity are all harms that arbitration can later remedy through damages or back pay, so they didn't justify emergency relief. That conclusion is consistent with earlier rulings the order cites involving Hawaiian Airlines and the University of Arizona.
Immigration status is the unresolved question
The one factor that gave the court real pause was immigration status. Four of the 26 plaintiffs hold employment-based visas sponsored by Meta, and Orrick's order notes they face a discretionary 60-day grace period to find new sponsored work, change status, or leave the country. They cannot work at all in the meantime. Calling that harm "more than speculation," Orrick said it could constitute the kind of irreparable injury that would support an injunction but said he didn't yet have enough information to rule on it. He's ordered Meta to file a declaration by July 23 explaining why those four plaintiffs specifically were selected for the RIF, with plaintiffs able to respond by August 6.
The plaintiffs' broader request for a preliminary injunction, which would last through the arbitration process rather than just a short-term TRO, remains pending. Meta must respond by August 10, and Orrick has set a hearing for August 24.
The case is being watched closely as what appears to be the first lawsuit against a major US tech company to directly challenge the use of AI in a layoff selection process, arriving roughly a month after a federal judge allowed a similar AI hiring-bias case against Workday to proceed, a ruling HRD covered in detail. Where Workday concerned algorithmic screening at the hiring stage, the Meta case tests whether behavioral-monitoring data feeding into performance and productivity scores can create a discriminatory outcome at the layoff stage, even without any intent to discriminate.
The ruling underscores how much weight documentation of human decision-making in a RIF can carry in court: it's the evidence that separates a "serious question" from a stronger claim. The case also lands amid a broader debate over where accountability sits when AI tools inform high-stakes employment decisions, a theme HRD has explored around the personal legal exposure AI tools can create for HR professionals and the gap between AI adoption and AI governance that many organizations are still working to close.
Any monitoring-based metric, whether keystrokes, token usage or general activity tracking, carries the same structural risk the Meta plaintiffs are alleging if it isn't adjusted for protected leave, regardless of whether a human or an algorithm makes the final call.
