Share with your CHRO
Twenty-six former Meta employees are suing the company over its May 2025 workforce reduction, which cut roughly 8,000 people, alleging that a suite of internal AI tools called Metamate and related performance-scoring systems disproportionately selected workers on parental or medical leave for termination. The lawsuit claims the AI ranking system penalized employees for exercising legally protected leave rights rather than excluding that period from their performance data. Meta denies the claims, insisting humans made all workforce decisions.
What this means for your business
Any organization running AI-assisted performance management sits directly in the blast radius of this lawsuit, whether or not it has a product named Metamate. The legally exposed position is specific and reproducible: a scoring model trained on activity signals like code commits, messages sent, or AI tool usage will mechanically undercount output for anyone on protected leave. The model doesn’t know it’s discriminating. It just scores what it sees, and courts are now being asked to treat that outcome as equivalent to intentional targeting.
Meta’s defense, that humans made the final calls, is the standard rebuttal in AI-assisted HR decisions, and it’s becoming legally fragile. The argument assumes a meaningful human review layer sits between the AI’s ranked list and the termination letter. But when a system surfaces a ranked list of 8,000 people to be cut and managers are working under time pressure, the AI’s ordering effectively is the decision. Regulators and plaintiffs’ attorneys increasingly understand this dynamic, which is why “a human approved it” stops being a shield the moment discovery reveals how much the ranked output drove the outcome.
The leading indicator worth watching is whether Meta’s internal documentation shows any explicit leave-exclusion logic in the scoring design. If it doesn’t, that absence becomes the case. For any CHRO whose team is piloting AI workforce analytics, the design question to answer before the next RIF (reduction in force) is whether your vendor’s model can be audited to confirm it excludes protected-status periods from its training window entirely, not just that a manager had final approval. That’s the difference between a defensible process and a discoverable liability.
Concept deep-dive: Protected-class data exclusion
In AI workforce scoring, protected-class data exclusion means deliberately removing performance signals generated during legally shielded periods, such as FMLA or parental leave, before a model ranks employees. Think of it like pausing a runner’s race clock during a mandatory rest stop. Without this exclusion, a model trained on continuous activity data treats absence as underperformance. It’s a required design choice, not a default behavior, and most off-the-shelf performance AI tools don’t implement it automatically.
Based on reporting from Meta accused of using biased AI targeting for mass layoffs, originally published 2026-07-14 13:18:00.
