On July 13, 2026, twenty-six current and former Meta employees filed a class-action complaint in federal court in Oakland, California. The claim: Meta's internal AI systems treated protected leave as a proxy for low output, then used those scores to steer layoff decisions that cut 8,000 workers in May 2026.
TL;DR: 26 Meta employees filed a federal class-action on July 13, 2026 alleging that Metamate, keystroke dashboards, and AI-native ratings scored workers on protected leave as low output and steered layoff decisions. Judge William Orrick denied an emergency injunction. Terminations were finalized July 22. Legal claims include FMLA, ADA, Pregnancy Discrimination Act, and Pregnant Workers Fairness Act. Meta says humans, not AI, made all decisions.
The case, styled Does 1-26 v. Meta Platforms, Inc., is the first federal lawsuit to directly challenge AI-assisted workforce reduction at a major tech company since the Workday litigation. It arrives at a moment when AI performance tooling has spread through every layer of tech company operations, and the legal infrastructure to govern it has not kept pace.
U.S. District Judge William Orrick of the Northern District of California denied the plaintiffs' emergency injunction request on July 17. Their terminations were finalized on July 22.
What Meta's AI Systems Measured
The complaint names five distinct data systems Meta used to rank employees before the May 2026 reduction in force:
Metamate is Meta's internal large language model. The complaint describes it as an employee-trained "second brain" that monitored internal communications, documents, and work product. The system generated content-quality and engagement scores from employee-produced material.
Second Brain is a related internal system also cited in the complaint, used to index and analyze employee-generated knowledge work and produce a parallel scoring signal.
Keystroke and activity monitoring captured keyboard inputs, screen activity, email usage, and browser history. Employees on leave generate zero input. The system had no logic to exclude leave periods from baseline calculations, so an eight-week absence registered identically to eight weeks of low productivity.
AI token-usage dashboards tracked how actively each employee used Meta's internal AI tools, measuring raw token consumption as a proxy for engagement with the company's AI strategy. Employees who were not at their workstations during leave periods consumed no tokens by definition.
AI-native ratings categorized every employee as "AI Native," "AI First," or "AI Enabled" based on their adoption of internal AI tools. An employee on twelve weeks of FMLA leave who returned to find that new tools had been rolled out in their absence faced an adoption gap they could not close before the scoring window closed, through no fault of their own.
All five systems fed into algorithmically generated performance rankings that managers reviewed before making final layoff selections. The complaint alleges that managers were not informed which employees' scores reflected leave-period absences versus actual underperformance. Human review was applied to AI-filtered information, not raw performance data.
Who Filed the Lawsuit
The 26 Does are not a random sample of the 8,000 employees who lost their jobs. The complaint identifies them by protected leave type:
Eight women had taken maternity or pregnancy-related leave within the past 24 months. Four men had taken parental leave. One woman had taken leave to care for a family member and later taken bereavement leave. The remainder had requested disability accommodations or taken medical leave for health conditions.
About half of the plaintiffs took caregiving or pregnancy leave in the 24 months before the May 2026 reductions. They span roles from engineers, scientists, and designers to researchers, managers, and directors, and are spread across six U.S. states and Washington, D.C.
None of them knew when they took leave that the preceding 24-month window would become the scoring period that determined whether they kept their jobs. The algorithm did not know either. It simply measured the gap.
The Legal Claims
The complaint names four statutes:
Family and Medical Leave Act (FMLA) prohibits using FMLA leave as a negative factor in employment decisions. Any scoring system that counts leave absences as low output without an exclusion flag is potentially using leave as a negative factor, even if that was not the intent of the system's designers.
Americans with Disabilities Act (ADA) prohibits discriminating against qualified individuals on the basis of disability. Employees who took medical leave related to a disability and then faced an AI score reflecting their leave-period absence are the core ADA plaintiffs in this case.
Pregnancy Discrimination Act (PDA) prohibits treating pregnancy, childbirth, or related medical conditions as a basis for adverse employment action. Eight of the 26 plaintiffs are PDA plaintiffs; their maternity leave appears in the scoring window as eight weeks of zero activity.
Pregnant Workers Fairness Act (PWFA) was enacted in 2023 and requires reasonable accommodations for pregnant workers. The complaint argues that using AI scores that failed to account for PWFA-protected accommodations violated the statute's anti-retaliation provisions.
Meta's Defense
Meta's statement through a spokesperson was direct: "The claims lack merit and are not based on facts. Workforce management and organizational decisions were and are made by people, not AI. We will defend ourselves vigorously."
The "humans made the decisions" framing is the standard defense in algorithmic employment litigation. It is also the defense that failed to get the Workday case dismissed at the pleading stage. Courts have found that when AI-generated rankings are the primary input that human decision-makers review, the human reviewer does not automatically insulate the AI tool from disparate impact analysis.
Meta's argument faces a specific evidentiary problem: if the complaint's allegations are accurate, managers were shown rankings without disclosure that certain scores reflected leave-period absences rather than actual work output. A human decision made on AI-filtered data, where the filter introduced the discrimination, is not the same as an unassisted human judgment.
The Invisible Worker Problem
The core technical failure here is not unusual or specific to Meta. It is the default state of most productivity measurement systems.
When an employee stops generating keystrokes, emails, Slack messages, code commits, and AI token consumption, the monitoring system records absence as zero activity. Zero activity scores as low productivity. Low productivity generates a low ranking. The low ranking reaches a manager who does not know the employee was absent because of a legal right rather than because of underperformance.
The fix is not architecturally complex: exclude leave periods from scoring windows before computing any percentile rank or calibration score. But it requires an explicit design decision. Engineers building productivity dashboards optimize for the median case. The median employee is not on FMLA. So the edge case never makes it into the design requirements document.
This is what makes the case significant beyond Meta. The invisible worker problem exists in every productivity monitoring system that does not have a leave exclusion flag built in from the start. That includes most of them, because the people who build these systems do not work in HR compliance and the people who work in HR compliance are not typically involved in systems design.
The Workday Parallel
This case will be litigated against the backdrop of the Workday litigation, where plaintiffs alleged that Workday's AI-assisted hiring tools screened out candidates based on race, age, and disability. The Workday case established that plaintiffs can pursue an AI tool vendor under disparate impact theory without proving discriminatory intent at the time the tool was designed.
The Meta case is the employment-side version with one critical difference: Meta is both the tool developer and the employer. There is no vendor to deflect to. Meta built Metamate, configured the scoring systems, set the 24-month scoring window, and applied the results to 8,000 termination decisions. The complaint argues that Meta is therefore fully responsible for every discriminatory output those systems produced.
The legal question courts will have to answer: does a company that builds and deploys AI performance tools internally assume the same liability as a vendor that sells similar tools to employers? If the answer is yes, every major technology company that built internal AI scoring systems during the 2022-2025 wave of internal AI deployment faces a similar exposure in any future reduction in force.
Five Governance Questions Before Any AI-Assisted Workforce Reduction
Whether you are a 20-person startup or a 100-person company, if any AI-assisted tool influences who gets selected for a reduction in force, these questions apply before execution:
1. Does your scoring window explicitly exclude protected leave periods? Pull one employee's leave record and cross-reference it against their performance score for the same window. If their score dropped during a leave period, you have a design flaw that creates legal risk. This check takes an afternoon and requires no new tooling.
2. Do your managers know which scores reflect leave absences? If the ranking dashboard does not flag leave-affected scores, the human reviewer is working with AI-filtered data that hides the protected-leave signal. The human review step cannot correct a discrimination that it cannot see.
3. Does your AI-adoption metric penalize employees who were absent during tool rollouts? An AI tool rolled out during someone's FMLA leave will show that employee as a non-adopter even if they became an active user the day they returned. That gap reflects timing, not performance. Rating employees on AI-tool adoption without accounting for absence creates the same exposure as Meta's AI-native rating system.
4. Have you run a disparate impact analysis on your reduction in force selection before executing it? Under Title VII and the ADA, disparate impact can be proven statistically without evidence of intent. Running a pre-layoff statistical analysis and documenting remediation steps is both a legal defense and a governance practice. Most small teams skip it, assuming the obligation only applies at scale. It does not.
5. Does your process include a step where someone asks who on this list was on protected leave in the last 24 months? It does not have to be automated. A human with an HR leave ledger can check 50 names in an afternoon. That check is what the 26 plaintiffs will argue Meta never performed. Not doing it is the governance failure discovery will surface.
The EEOC guidance on AI in employment decisions covers the federal framework for this analysis. NYC Local Law 144 requires bias audits for automated employment decision tools used in New York City. The FCRA disclosure requirements cover a parallel obligation when third-party data sources feed into employment decisions.
What the July 22 Deadline Settled and What It Did Not
The 26 plaintiffs sought an emergency injunction to halt their terminations. Judge Orrick denied it. The terminations were finalized on July 22.
That denial does not resolve the underlying case. The standard for emergency injunctive relief is deliberately high: a strong likelihood of success on the merits, irreparable harm not remedied by damages, and a favorable balance of equities. Employment terminations, however damaging to individual plaintiffs, are typically remedied through back pay and reinstatement if the plaintiff prevails at trial. Courts regularly deny emergency relief in employment cases while allowing the underlying claim to proceed.
The class-action now moves into discovery. The most consequential documents Meta will be required to produce are the algorithm specifications for Metamate's performance scoring, the token-usage dashboard methodology, the AI-native rating rubric, and any internal communications discussing whether protected leave should be excluded from scoring windows during system design.
If those documents show that engineers or product managers discussed the leave exclusion problem and decided not to implement it as a design feature, the litigation becomes significantly more difficult for Meta to defend. If the documents show the issue was never raised, the claim is negligent design rather than willful discrimination. Both paths lead to liability; they differ in exposure and remedies.
The Broader Signal for AI-Powered Workforce Management
Meta is not the only technology company that built internal AI productivity monitoring systems during the 2022-2025 period. It is the first to face a federal class-action over how those systems were used in a major reduction in force.
The significance is that discovery in Does 1-26 v. Meta Platforms will surface the algorithm specifications. Those specifications, once in the public record, will become a template that plaintiffs' attorneys use when evaluating claims at other companies with similar internal tooling. The invisible-worker problem will be examined in detail and in public for the first time in a federal court proceeding.
HR teams that have not audited their AI performance systems for leave exclusion logic now have a concrete reference point: a filed complaint, a named set of systems, and a clear factual pattern. The question their legal teams are about to start asking is not whether the system was intended to discriminate. It is whether the system, as built, produced discriminatory output in its scoring window.
Intent is not a field in the algorithm. Output is.
For teams using any form of AI-assisted performance tracking, the meta MCI employee monitoring governance analysis covers the monitoring side of the same internal tooling. The AI hiring compliance state matrix covers where state-level obligations apply to automated employment decisions beyond federal law.
Related Reading
- Workday AI Lawsuit: What HR Teams Must Know Before Using AI Screening
- EEOC AI Hiring Guidance: Employer Checklist for AI Employment Tools
- FCRA and AI Hiring: Disclosure Requirements for Automated Screening
- NYC Local Law 144: AI Bias Audit Requirements for Employers
- Meta MCI: What the Employee Keystroke Tracking Leak Means for Your Monitoring Policy
- AI Hiring Tool Compliance: US State-by-State Matrix 2026
- AI Acceptable Use Policy Template for Small Teams
- AI Vendor Due Diligence Checklist 2026
- AI Models 2x More Likely to Refuse Criticizing Repressive Regimes
