A federal judge in San Francisco spent Monday morning telling 26 anonymous former Meta employees that the hardest part of suing an algorithm is proving the algorithm was there at all. "I have a record I have to deal with and the record at the moment does not persuade me of the merits," U.S. District Judge William H. Orrick said at an August 24 hearing on the workers' motion for a preliminary injunction. It was the second time in five weeks he declined relief, and the clearest signal yet that the first major test of AI-driven layoffs may never be decided on the AI question at all.
The case is Does 1 Through 26 v. Meta Platforms, Inc., No. 3:26-cv-07122-WHO, filed July 13 in the Northern District of California. All 26 proceed anonymously; each took, requested, or was approved for protected leave -- medical, parental, pregnancy-related, or disability accommodation -- within the 24 months before Meta announced in April that it would cut roughly 8,000 jobs, about 10% of its global workforce. Notices went out in May. The complaint runs to 21 causes of action spanning the ADA, the FMLA, Title VII, the Pregnancy Discrimination Act, the Pregnant Workers Fairness Act, and state statutes in California, Washington and New York.
The theory is disparate impact, not animus. Plaintiffs allege Meta used what the complaint calls a "constellation" of internal AI systems to score, rank and select employees for the reduction in force: Metamate, an internal LLM assistant; "second brain" agents trained on employees' own documents and communications; productivity scoring drawn from keystrokes, screen content, mouse activity, browser history and messaging pulled from company devices; and dashboards tracking AI-token consumption. None of those meters, plaintiffs argue, can be moved by someone legitimately not at work; the systems were not paused for approved leave, so scores fell, and the fall fed the list.
Meta's denial is categorical. Opposing the earlier restraining order, it submitted a declaration from Linh Doan, director of HR business partner enablement, saying selections were made by human leaders applying documented, neutral criteria -- job level, performance history, tenure, location, function, specialized skills and reporting-structure efficiency -- and that no plaintiff was chosen because of leave status or disability. Meta attorney Erin Connell put it flatly: "This whole matter started with allegations that Meta used AI to determine which employees to include in the reduction in force. There is no evidence of that. That did not happen. That remains true."
Orrick has accepted neither account. His July 17 order denying the TRO found plaintiffs had raised "serious questions going to the merits" -- the lower rung of the Ninth Circuit sliding-scale test -- but not a likelihood of success, holding that lost pay, health coverage and unvested equity are harms arbitration can remedy later. Separations proceeded July 22. On Monday he said the plaintiffs' evidence "raised some potential questions about Meta's categorical denial of any impact of AI in the termination process," but still did not warrant an injunction. The closest call is immigration: four plaintiffs hold Meta-sponsored visas and face a 60-day window to find new sponsorship or leave the country. "I think the issue could be irreparable," Orrick said, "but again, I don't think the merits are close enough to warrant relief."
Andrea Mazingo of Lumen Law Center, arguing for the workers, said "the relief we're requesting is modest and severable" -- principally an independent audit of the selection process. Orrick called the dispute "an unusual, or a new sort of issue" and said an opinion would follow shortly. Per MLex, he indicated the claims are likely better suited for arbitration, where plaintiffs can still pursue that audit.
Why It Matters
Strip away the injunction fight and one sentence from the bench is the whole case: Orrick said it turns on "whether what plaintiffs think happened, happened." That is a discovery question, not a merits question -- and it is where this entire line of litigation gets decided.
Disparate-impact law was built for records employers already keep: test scores, cutoff thresholds, applicant flow data. It maps badly onto a scoring pipeline whose inputs are telemetry and whose weights are a model. A plaintiff cannot show adverse impact without the model, its features and its outputs; the employer alone holds all three; and a plausible claim is the price of discovery. That circle is the defense.
Mobley v. Workday, the ADEA collective action before Judge Rita Lin in the same district, shows how it closes. Workers cleared the pleading stage and won court-authorized collective notice in February 2026. Then, on May 29, 2026, Magistrate Judge Laurel Beeler denied their motion to compel Workday's bias-testing data, holding it protected by attorney-client privilege. Workday had to produce its own EEO-1 and OFCCP filings -- the conventional records -- but the algorithmic self-examination stayed sealed. Run your bias testing under counsel, and the most probative evidence about your model may become the least discoverable.
Statutes have not solved this. California's FEHA automated-decision-system regulations took effect October 1, 2025, requiring four-year retention of selection criteria, outputs and audit findings; plaintiffs invoke them here. But retention is not disclosure. New York City's Local Law 144 mandates annual bias audits, yet a Comptroller audit found enforcement ineffective; Colorado replaced its AI Act with a narrower transparency regime effective January 1, 2027. And arbitration carries its own limited discovery rules.
What to Watch
Orrick's written order is imminent and will likely deny the injunction; read it for the evidentiary standard it sets, not the outcome. The real signal comes next: whether these claims land in arbitration, what discovery an arbitrator permits into Metamate and the productivity-scoring inputs, and whether Meta asserts the same privilege that shielded Workday's bias testing. Watch, too, the four visa-dependent plaintiffs -- the one place Orrick conceded the harm may be irreparable. Nothing here has been proven. But the question of who gets to look inside the model is being answered right now, largely out of public view.
“I have a record I have to deal with and the record at the moment does not persuade me of the merits.”— William H. Orrick, U.S. District Judge, Northern District of California