Twenty-six current Meta employees have sued the company, claiming that a layoff selection process combining internal AI tools and activity metrics put people who took medical, family, or pregnancy-related leave at a disadvantage. The plaintiffs argue that output volume and AI usage, which cannot be built up while on leave, were converted into lower evaluations. Meta flatly denies the claims, saying "decisions about staffing and organization were made by humans." The dispute is not over whether AI alone decided who would be fired. It is over which data moved candidates up or down the ranking, and how far humans were able to overturn it.
The 26 plaintiffs want to stop their terminations from becoming final
The complaint was filed on July 13, 2026, in the U.S. District Court for the Northern District of California. The 26 plaintiffs are anonymous and were among those targeted in the layoffs Meta began notifying on May 20, which affect about 8,000 people, roughly 10% of its workforce. Meta reported 77,986 employees as of the end of March. Most plaintiffs are scheduled to leave on July 22; the complaint lists around August 20 for the New York plaintiffs and around July 31 for one plaintiff in Washington State.
The 26 say that in the 24 months before the selection, they took or requested protected leave, were prevented from taking leave, or requested or received accommodations for a disability. Ten of the plaintiffs allege pregnancy or sex discrimination. According to the Associated Press, eight women had taken childbirth- or pregnancy-related leave and four men had taken parental leave. One woman who took family caregiving leave and bereavement leave is also included.
What the plaintiffs have asked the court for is not a collective ruling on all 26 claims as a class action. Because Meta's employment contracts contain individual arbitration and class-action waiver clauses, each plaintiff plans to contest the merits in arbitration. Before that, they have moved for a preliminary injunction to maintain their employment and pay, and to keep medical coverage, unvested stock compensation, and leave status unchanged. They are also demanding an independent audit of the selection process. Their reasoning for seeking relief now is that monetary compensation after separation could not undo the loss of health insurance during treatment, expiring leave entitlements, or effects on immigration status.
The layoffs coincided with Meta's rapid expansion of AI investment. In its April 29 earnings report, the company projected 2026 capital expenditures of $125 billion to $145 billion, saying the spending would support AI and its core business, up sharply from the 2025 actual of $72.22 billion. But the spending figure does not prove that discriminatory selection took place. It only shows the business backdrop: a layoff of around 8,000 people and internal AI adoption advancing at the same time.
The alleged process: a mix of AI usage and output volume
The selection process the plaintiffs describe is not a single "firing AI." The complaint names the internal assistant "Metamate" and "second brain"-style agents that employees train themselves. These are followed, the plaintiffs say, by device activity monitoring, a dashboard of AI token usage, and AI-assisted performance reviews, which together scored employees and fed them into the layoff candidate pool.
The inputs the complaint alleges went into the selection are specific. First are performance ratings for the most recent 12 months and the results of end-of-2025 calibration. Added to these, the plaintiffs say, were an "AI-native" rating reflecting how much employees had integrated AI into their work, output volume, and code commit counts, with managerial sponsorship and proximity to priority roadmaps also taken into account. One plaintiff, an employee at the director level, stated under oath that an internal program called "Checkpoint" announced that AI adoption would be a key evaluation criterion, and that the usage dashboard had no mechanism to distinguish being on leave from low engagement.
This is where the measurement problem lies. If code commits or AI token volume are counted against the same 12-month denominator, someone who took three months off can only accumulate numbers for the nine months they worked. Unless the leave period is excluded from measurement, normalized by days actually worked, or routed to individual review, lawful absence is recorded as low productivity. People who adjusted their input methods or working hours because of a disability face the same problem if compared on the volume of device activity.
However, what was actually fed into the selection, and how much each item moved the rankings, has not been established. Much of the technical core of the complaint is stated "on information and belief," and the specifications of the selection model and the distribution across all those targeted are not currently public. The plaintiffs are seeking an independent audit because they cannot yet verify the inputs, weights, outputs, and managers' override history themselves.
Human approval can still fall under rules on automated decisions
Meta's rebuttal is straightforward. It told the AP that the claims are not grounded in fact and that decisions about staffing and organization were made by humans, not AI. But the legal boundary is not determined solely by who pressed the final button.
In California, regulations on automated-decision systems in employment took effect on October 1, 2025. The state's Civil Rights Council covered computational processes that screen or evaluate applicants or employees, from classification to recommendation to decision-making, including those that merely assist human decision-making. A workflow in which AI ranks candidates and a manager approves the names is not automatically excluded from that definition.
The regulations make clear that an automated-decision system that disadvantages applicants or employees because of protected characteristics such as sex, race, or disability can violate state law. Whether proxies for protected characteristics were used, and the scope and frequency of testing conducted to curb discrimination, can also be factors. Employment records must also include automated-decision data and be retained for at least four years. The plaintiffs here have demanded preservation of the full set of related materials, including selection data, models, and decision logs, because without this evidence the contributions of humans and the system cannot be separated.
So "humans decided" and "AI assisted the selection" can both be true. Checking what inputs and weights were used in the selection, and how leave periods were excluded, gets closer to the reality than the label does. The share of recommendations managers overturned, and the selection rates of employees who took leave versus those who did not, are also material for testing outcomes and causation.
Leave doesn't exempt anyone from layoffs, but it can't count against them
The federal Family and Medical Leave Act (FMLA) prohibits using qualifying leave as a negative factor in employment decisions. It is not a provision that categorically exempts employees on leave from workforce reductions. If an employer can show the employee would have been selected for the same reasons even without taking leave, including that employee in a reduction can be permissible. In this dispute, the question is whether output volume and AI usage acted as a proxy for leave, even if leave itself was never entered as an input.
The Americans with Disabilities Act (ADA) likewise does not treat evaluations made through software or AI as an exception. In 2022, the Equal Employment Opportunity Commission (EEOC) and the Department of Justice warned that employment decisions using AI require procedures for providing reasonable accommodations, and that without appropriate safeguards, even people with disabilities who can perform the job may be screened out. A design that measures device activity or output volume needs a way to avoid mistaking differences caused by workplace accommodations for lack of ability.
For the sex-related claims, besides intentional discrimination, "disparate impact" is at issue. 42 U.S.C. § 2000e-2(k) provides that even a facially neutral employment practice can be unlawful if it produces results skewed against a protected group such as one defined by sex and the employer cannot show it is job-related and consistent with business necessity. The plaintiffs argue that a mechanism counting leave as reduced performance would fall most heavily on women, who take more leave for pregnancy, childcare, and caregiving.
The regulatory environment is not monolithic. In Executive Order 14281 of April 2025, the Trump administration ordered federal agencies not to prioritize enforcement based on disparate impact and called for a review of ongoing investigations and lawsuits. Even so, the text of § 2000e-2(k), enacted by Congress, remains. California's automated-decision regulations and state anti-discrimination laws also remain in force, and the route for individual plaintiffs to bring civil claims has not been closed.
By July 22, the question is whether the process can be audited
This lawsuit does not, on the experiences of 26 people alone, prove that Meta's selection as a whole was statistically discriminatory. It has not been shown how many of all those targeted were on leave or at what rate they were placed in the layoff pool. For each plaintiff, it would also need to be tested whether they would still have been selected on an evaluation that excluded leave. Meta's formal response and evidence submissions are still to come.
Still, the plaintiffs' demands make the method of verification clear: an independent auditor examines the selection's inputs, weights, and outputs and checks whether leave, disability accommodations, or proxies for them were used. Scores are then recalculated with leave periods neutralized, to see whether a reasonable basis for selecting each plaintiff remains.
For companies, the issue cannot be settled by dropping the name "AI." Whether it is a spreadsheet score or a machine-learning ranking, anything that influences who is sent to the layoff candidate list requires that the validity of the inputs and the handling of exceptions can be explained. With the plaintiffs' July 22 separation date approaching, the near-term question is how the court assesses the urgency of the motion and the risk of irreparable harm.