September 29, 2026
What Meta’s lawsuit means for employers that use automated scoring
Twenty-six current and former Meta employees filed a lawsuit in July, alleging that the company’s metrics penalized workers for absences from protected leave or accommodations for disabilities. The plaintiffs claim that Meta “used a constellation of internal artificial intelligence systems,” including one that monitors keystrokes and computer activity, “to score, rank and select employees for inclusion on the list” for its May 2026 layoffs. According to the plaintiffs, the systems kept recording reduced activity while employees were on protected leave, so workers who took leave were disproportionately selected. Meta has denied the accusations, stating that workforce decisions “were and are made by people, not AI,” and submitted a declaration that no plaintiff was selected based on leave status, disability, or any other protected characteristic.
The complaint
The plaintiffs’ theory is that scores built on activity data will penalize anyone who is absent, including those on protected leave. Meta says humans, not scores, chose who to lay off based on job profile or level, performance ratings, tenure, location, job function, specialized skills, and reducing management layers. Further, Meta states there was “no AI-assisted ‘scoring’ or ‘ranking’ related to employee performance.” This case is defined by two questions: (1) did Meta’s metrics penalize protected leave or disability, and (2) were those metrics what determined who was laid off?
Regarding the first question: federal law is clear that employers cannot use FMLA leave as a negative factor in employment decisions, and the ADA prohibits treating qualified employees differently because of a disability, including layoff selection. The second question changes how the plaintiffs prove they were discriminated against. The plaintiffs, to show that leave or disability was a factor in their selection for a layoff, asked the court to order a “leave-neutralized” re-computation of their metrics. Meta called their request “discovery disguised as injunctive relief” and said it belongs with the arbitrator.
Showing that humans made the decisions would not end the case. Managers working with biased scores can still be unlawful. Under the ADA, criteria that tend to screen out people with disabilities are unlawful unless the employer shows otherwise, and the plaintiffs raise a similar pregnancy-based theory. A “leave-neutralized” re-computation also would not resolve a disability claim, because reduced output caused by a disability is not the same as time away.
The case is at an early stage. Judge William Orrick denied the employees’ request for a temporary restraining order to halt their layoffs in a July 17 order, finding that most of their losses could be remedied with money later. He found, however, that the plaintiffs had raised “serious questions going to the merits” of their claims. He added that he may revisit his ruling based on evidence “regarding whether and how AI was used.” On the merits, the court found Meta’s declarations that it did not use AI “unequivocal,” noting that the plaintiffs, who were not “in the rooms” where the decisions were made, lacked the evidence to rebut them. Because four plaintiffs on Meta-sponsored visas could face irreparable immigration consequences, he also ordered Meta to explain “how and why” those four were selected. The merits of the discrimination and leave claims will be decided in individual, confidential arbitration; the court proceeding concerns only whether the plaintiffs are reinstated in the meantime. All their terminations have since taken effect. At an August 24 hearing on their preliminary injunction motion, Judge Orrick said the plaintiffs’ new evidence “raised some potential questions about Meta’s categorical denial of any impact of AI in the termination process” but did not persuade him that an injunction was warranted, and he indicated a written ruling would follow.
If the court did rule against Meta in the interim, it would put the plaintiffs back in their jobs and benefits until arbitration ends, with the outcome of arbitration determining whether other remedies, such as damages, would follow. In arbitration, Meta can argue that leave was not factored into its selection criteria, or that the plaintiffs would have been selected even if they had not taken leave. For employees laid off while on leave, however, the employer bears the burden of proving the second point, according to regulations from the Department of Labor.
How ADMT laws would change the case
This case offers an early look into how automated decision-making technology (ADMT) laws could apply to a dispute like this. California’s civil rights regulations on automated-decision systems, in effect since October 1, 2025, cover any “computational process that makes a decision or facilitates human decision making regarding an employment benefit,” regardless of the technology1 and even when a human makes the final call. For Meta, two effects would follow: (1) it would have to retain any selection criteria and system outputs behind the layoffs for four years, and (2) the absence of anti-bias testing could be used as evidence against it. For plaintiffs with California disability or accommodation claims, these rules apply directly; how far they reach leave-based claims is less clear. Because these regulations were in effect before the layoffs, they are already part of this case: the plaintiffs’ preliminary injunction motion cites them, along with California’s disability and family-leave laws.
Two more laws take effect on January 1, 2027, and both are defined by the case’s second question: how much impact the score had on the layoff decision. (Our overview of the California and Colorado laws covers them in detail.) Under Colorado, any computed score that “materially influences” a decision, whether AI produced it or a human signed off on it, is ADMT2. While none of the plaintiffs appears to work in Colorado, an employee there would be owed notice when the scoring tool was first used (likely long before the layoffs), an explanation of the score’s role in the layoffs, and a chance to correct inaccurate data and request a human review. In this case, the right to correct does not apply to the scores, but to the data that was used to determine the scores. A plaintiff in a similar case could argue that recording a leave period as zero activity is inaccurate. California’s ADMT regulations apply when an ADMT’s output is used without meaningful review by a qualified human. If Meta’s managers approved the rankings without that level of review, Meta would have owed a pre-use notice, an opt-out or human appeal, and, on request, an explanation of the tool’s logic and its effect on each employee. Even if the review was genuine, Meta would still have had to complete a risk assessment, if, as alleged, its scores were built by systematically monitoring employees’ work.
The 2027 laws are enforced by regulators, not by employee lawsuits, but the civil rights rules can support an employee’s own claim, and the records all three regimes require could be sought as evidence in arbitration. For plaintiffs in a case like this one, those records would show what effect, if any, protected leave had on an employee’s score, and how much, if at all, that score drove the decision to lay the employee off. At the restraining-order stage, the court had to take Meta at its word. Under these laws, the answers would already be on records Meta was required to create before anyone sued. Chicory helps organizations stay in compliance with automated decision-making laws and actively monitor automated systems, AI or not, for bias, keeping your employment practices transparent and legally sound.
- Basic tools such as spreadsheets are excluded unless they make the decision.
- “Materially influenced” is to be defined by the Colorado Attorney General.