Did Meta's layoff process penalize protected leave? Public records cannot calculate the selection rates

The public record does not establish that protected leaveprotected leaveProtected leave is qualifying time away from work covered by a law that grants specified employment protections for family, medical or related needs. [Department of Labor](https://www.dol.gov/agencies/whd/fmla/law) increased a Meta employee's probability of being selected for termination. It establishes that 26 selected workers with protected-leave or disability-accommodation histories filed a lawsuit and that they attribute their selections to an AI-assisted process. Meta disputes that account. [ComplaintComplaintA complaint is the document that begins a civil lawsuit and states the plaintiffs' allegations and requested relief.](https://www.courthousenews.com/wp-content/uploads/2026/07/meta-employee-complaint.pdf) [Associated Press](https://apnews.com/article/meta-lawsuit-workers-target-ai-layoffs-leave-019fb9c7fdc09167e91547546bce5be8)

That distinction is not legalistic hair-splitting. A rate needs a selected count and the full group from which that count came. Knowing that 26 passengers missed a train cannot reveal the miss rate unless we also know how many passengers tried to board.

The public evidence does support a narrower finding: the measurement failure alleged by the plaintiffs is mathematically possible, legally consequential and testable with records Meta should be able to preserve. It is not yet proved.

Meta announced on April 23 that it would eliminate about 8,000 jobs, which it described as about 10% of its workforce. The company said the changes were intended to improve efficiency and permit new investment as it increased spending on artificial intelligence infrastructure and specialist employees. [Associated Press](https://apnews.com/article/microsoft-voluntary-buyouts-ai-224eee4489cbc227244558ff02f5919a)

The complaint says Meta began notifying selected workers on May 20. Six California WARN entries later listed 3,270 Meta positions scheduled for separation on July 22 across Fremont, Burlingame, San Francisco, Playa Vista, Menlo Park and Sunnyvale. Those records establish locations, dates and position counts. They contain no leave histories, accommodation records, performance scores or explanations for individual selections. [Complaint](https://www.courthousenews.com/wp-content/uploads/2026/07/meta-employee-complaint.pdf) [California EDD](https://edd.ca.gov/siteassets/files/jobs_and_training/warn/warn-report-for-7-1-25-to-6-30-26.pdf)

On July 13, the 26 workers asked a federal court to preserve their employment while they pursue claims principally through individual arbitrationarbitrationArbitration is a private process in which a neutral arbitrator, rather than a court, decides a dispute.. AP reported the following day that all 26 remained employed and that the first separations were scheduled for July 22. [Courthouse News](https://courthousenews.com/meta-employees-sue-over-use-of-ai-in-workforce-reduction/) [Associated Press](https://apnews.com/article/meta-lawsuit-workers-target-ai-layoffs-leave-019fb9c7fdc09167e91547546bce5be8)

The complaint alleges that Meta used a collection of internal systems and data sources, including its Metamate assistant, employee-trained AI agents, computer-activity measures, AI-token-usage dashboards and algorithmically assisted performance rankings, to help assemble the termination list. It says some of those measures could not accumulate while a worker was away on protected leave or producing less work because of a disability. [Complaint](https://www.courthousenews.com/wp-content/uploads/2026/07/meta-employee-complaint.pdf) [Courthouse News](https://courthousenews.com/meta-employees-sue-over-use-of-ai-in-workforce-reduction/)

The complaint further says all 26 plaintiffs had taken, requested or been approved for protected leave during the preceding 24 months, while AP reports that each plaintiffplaintiffA plaintiff is a person who brings a claim in court. either took protected leave or requested or received a disability accommodation. These descriptions establish the plaintiffs' alleged histories, not how common those histories were among all selected or retained workers. [Courthouse News](https://courthousenews.com/meta-employees-sue-over-use-of-ai-in-workforce-reduction/) [Associated Press](https://apnews.com/article/meta-lawsuit-workers-target-ai-layoffs-leave-019fb9c7fdc09167e91547546bce5be8)

Meta told AP that the claims lack merit and that its workforce and organizational decisions were made by people rather than AI. That response disputes the plaintiffs' central description, but it does not publicly disclose what information human decision-makers saw, how preliminary rankings were produced or how often managers changed them. [Associated Press](https://apnews.com/article/meta-lawsuit-workers-target-ai-layoffs-leave-019fb9c7fdc09167e91547546bce5be8)

The plaintiffs' strongest theory does not require a futuristic machine that independently chooses whom to fire. It requires only a measure that counts total activity over a fixed calendar period without properly accounting for days when a worker was not expected to work.

Imagine two employees who produce at the same pace whenever they are working. If one takes approved leave, that employee will record less total output across the calendar even though output per active workday is identical. The total is arithmetically correct, but it answers a different question from how productively each person worked while present.

NIST warns that converting complex human behavior into measurable quantities can remove context needed to understand an AI system's effects. In this case, the missing context would be whether a low activity total reflected weak work, fewer expected workdays or some combination of the two. [NIST](https://airc.nist.gov/airmf-resources/airmf/appendices/app-c-ai-risk-management-and-human-ai-interaction/)

The complaint alleges that Meta failed to neutralize protected absence before using disputed measures. The public record does not show the actual formulas, whether they used calendar days or active workdays, whether scores paused during leave, or whether the measures entered the selection process at all. [Complaint](https://www.courthousenews.com/wp-content/uploads/2026/07/meta-employee-complaint.pdf)

The alleged error could exist in an AI system, a dashboard or an ordinary spreadsheet. The decisive issue is how absence entered the measurement.

The 26 plaintiffs are people who were selected and then challenged their selections. Their presence demonstrates that selected workers with leave or accommodation histories exist. It does not show the proportion of all workers with those histories who were selected.

Investigators need two complete fractions: selected workers with the relevant history divided by all workers with that history who were considered, and selected workers without that history divided by all other workers considered. The EEOC describes a selection rateselection rateA selection rate is the number selected from a group divided by the number considered in that group. [EEOC](https://www.eeoc.gov/laws/guidance/questions-and-answers-clarify-and-provide-common-interpretation-uniform-guidelines) as the number selected from a group divided by the number considered in that group. [EEOC](https://www.eeoc.gov/laws/guidance/questions-and-answers-clarify-and-provide-common-interpretation-uniform-guidelines)

• All leave-takers and accommodation recipients considered in each actual decision unit.
• Comparable workers without those histories who were considered in the same units.
• Retained workers in both groups.
• Workers whose positions disappeared without an individual performance comparison.
• The complete selected counts, including selected workers who did not join this lawsuit.

Even complete raw selection rates would describe an association, not automatically identify a cause. Leave histories might differ across teams, roles, locations or projects that faced different levels of restructuring. Those factors could create or hide a company-wide gap.

The California WARN data illustrate this problem. They show where 3,270 affected positions were located, but not whether workers at those sites were compared with one another, selected through role elimination or ranked individually. Using the WARN total as a denominatordenominatorThe denominator is the full eligible group from which the selected count came, such as all leave-takers considered in the same decision unit. [EEOC](https://www.eeoc.gov/laws/guidance/questions-and-answers-clarify-and-provide-common-interpretation-uniform-guidelines) for the plaintiffs would therefore mix different populations and answer the wrong question. [California EDD](https://edd.ca.gov/siteassets/files/jobs_and_training/warn/warn-report-for-7-1-25-to-6-30-26.pdf)

The Department of Labor says an employer cannot punish a worker for using FMLA leave and cannot use that leave as a negative factor in an employment action. Its guidance specifically says protected leave cannot generate negative attendance points or remove positive points. [Department of Labor](https://www.dol.gov/agencies/whd/fact-sheets/28a-fmla-employee-protections) [29 CFR 825.220](https://www.ecfr.gov/current/title-29/part-825/section-825.220)

EEOC guidance makes the performance-measurement issue unusually concrete. It says a worker who misses several weeks for medical reasons should not receive a poor rating for failing to meet a production quota during the absence. Separate guidance says an employer may evaluate work actually produced while a person was present and may consider postponing an evaluation when substantial leave affects total productivity. [EEOC performance evaluations](https://www.eeoc.gov/employers/small-business/5-im-conducting-performance-evaluations) [EEOC disability guidance](https://www.eeoc.gov/laws/guidance/applying-performance-and-conduct-standards-employees-disabilities)

The ADA does not require employers to abandon legitimate production standards. It can, however, prohibit a selection procedure that screens out qualified people with disabilities unless the procedure is job-related and consistent with business necessity, and it can require a reasonable accommodationreasonable accommodationA reasonable accommodation is a workplace change that enables a qualified person with a disability to work or receive an equal employment opportunity unless it would cause undue hardship. [EEOC](https://www.eeoc.gov/laws/guidance/enforcement-guidance-reasonable-accommodation-and-undue-hardship-under-ada) in how a procedure is administered. [EEOC](https://www.eeoc.gov/laws/guidance/employment-tests-and-selection-procedures)

Protected leave is not immunity from a layoff that would have happened anyway. Under the FMLA regulation governing reinstatement, a worker has no greater right to continued employment than if continuously employed. If an employer lays off a worker during FMLA leave, the employer bears the burden of showing that the worker would have been laid off regardless. [29 CFR 825.216](https://www.ecfr.gov/current/title-29/part-825/section-825.216)

These rules explain why the alleged calculation matters, but they do not decide this case. No court has found that Meta used protected leave as a negative factor, that a disputed system caused any selection or that Meta violated the FMLA, ADA or another law.

The central allegation can be tested, but only if the analysis follows Meta's real decision structure rather than treating the company as one undifferentiated pool.

• The complete consideration pool and each worker's final selection outcome.
• The team, role, level, location and other boundaries within which workers were compared.
• Protected-leave dates, expected working days and relevant accommodation records.
• Earlier performance information collected before the disputed leave period.
• The formula, inputs, weights and version history for every disputed score.
• Preliminary rankings, final rankings, organizational elimination criteria and human override records.

First, analysts should reproduce every disputed score using the formula shown to decision-makers. They should then calculate an alternative that removes approved absence from the time being measured. If a score changes materially after this adjustment, that would show that absence affected the measure, not that the measure caused selection.

Second, analysts should compare workers within the same decision units and match people with similar roles, levels, locations, tenure and earlier performance. A regressionregressionRegression is a statistical method used to estimate how an outcome varies with one or more recorded factors. [NIST](https://www.itl.nist.gov/div898/handbook/quantgal.htm) model could then estimate the remaining selection difference while accounting for recorded confounders. Confidence intervals should show how uncertain those estimates are. [NIST](https://www.itl.nist.gov/div898/handbook/prc/section1/prc14.htm)

Third, the analysis should examine the chain from leave to score to decision. Mediation analysisMediation analysisMediation analysis examines whether one factor affects an intermediate measure that then affects the final outcome. [National Library of Medicine](https://pmc.ncbi.nlm.nih.gov/articles/PMC12206152/) asks whether leave changed an intermediate measure and whether that measure then helped change the outcome. Here, the theory gains support only if protected absence predicts a lower disputed score, that score predicts selection among comparable workers, and correcting the score meaningfully weakens the leave-selection relationship. [National Library of Medicine](https://pmc.ncbi.nlm.nih.gov/articles/PMC12206152/)

Finally, investigators should compare the preliminary and final lists. Frequent, documented and substantive human changes would support Meta's contention that people exercised independent judgment. A near-automatic approval process would make the upstream rankings more consequential. Either conclusion requires records, not labels.

Several explanations fit at least part of the public evidence. The available records cannot yet distinguish among them.

• Leave-sensitive measurement: unadjusted activity or output totals made protected absence look like underperformance, and those measures influenced selection. This is the plaintiffs' central allegation. [Complaint](https://www.courthousenews.com/wp-content/uploads/2026/07/meta-employee-complaint.pdf)
• Role or project elimination: Meta removed organizational units that happened to contain more workers with leave or accommodation histories, without using those histories to choose individuals. Meta cited efficiency and investment priorities when announcing the broader reduction. [Associated Press](https://apnews.com/article/microsoft-voluntary-buyouts-ai-224eee4489cbc227244558ff02f5919a)
• Independent human judgment: managers used job requirements, conventional performance information and organizational needs rather than AI-generated rankings. This is consistent with Meta's public response but requires internal instructions and decision records to verify. [Associated Press](https://apnews.com/article/meta-lawsuit-workers-target-ai-layoffs-leave-019fb9c7fdc09167e91547546bce5be8)
• Mixed processes: different teams used different inputs or levels of human review, producing individual disputes without one company-wide mechanism.

The most informative comparison is therefore within the units where decisions were actually made. If an apparent leave gap disappears among workers in the same role, team and elimination category, restructuring becomes a stronger explanation. If the gap remains and follows score changes during leave, the measurement theory becomes stronger.

• Evidence supporting the alleged mechanism would include a score decline during protected absence, a relationship between that score and selection within comparable groups, and few meaningful human overrides.
• Evidence weakening the mechanism would include scores paused or adjusted during leave, proof that the disputed measures were not shown to decision-makers, or no adjusted relationship between the scores and selection.
• Evidence supporting a company-wide disparity would require complete selection rates for properly defined groups, with enough information to assess uncertainty and organizational differences.
• Evidence against a company-wide disparity would not necessarily resolve individual retaliation, interference or accommodation claims, which can depend on facts specific to one worker.

The immediate court dispute concerns temporary protection before the scheduled separations while the workers pursue their underlying claims. Whatever the court decides about temporary relief, that procedural decision should not be mistaken for a final statistical or legal finding about Meta's selection process. [Complaint](https://www.courthousenews.com/wp-content/uploads/2026/07/meta-employee-complaint.pdf) [Law360](https://www.law360.com/employment-authority/wage-hour/articles/2500645/meta-employees-say-ai-tainted-layoffs-should-be-blocked)

The complaint identifies a credible way that productivity measurement could turn protected absence into apparent underperformance. Federal guidance confirms that the distinction between work missed during protected leave and work performed while present can matter. [Department of Labor](https://www.dol.gov/agencies/whd/fact-sheets/28a-fmla-employee-protections) [EEOC](https://www.eeoc.gov/employers/small-business/5-im-conducting-performance-evaluations)

But the public evidence does not show Meta's formula, establish that the disputed measures reached decision-makers or reveal the selection rates for comparable workers. Even the apparently known table cell is only a lower bound from the plaintiffs, not a complete company count.

The public record establishes a mechanism worth investigating. It does not establish that protected leave increased the company-wide probability of selection.

The honest answer is therefore: possibly, but not proved. Resolving the question requires reconstructing the path from expected workdays to scores, from scores to preliminary rankings, and from those rankings to the final human decisions.