Meta used AI to tag workers who took leave to be laid off, lawsuit claims
According to a 71‑page federal complaint filed in the Northern District of California, a Meta scientist says she received a layoff notice two days before giving birth, and the plaintiffs allege internal AI scores and device telemetry, not manager judgment, produced the layoff list. For executives, the combination of employee surveillance, automated scoring and opaque governance the complaint describes is a legal and reputational risk any company using employee telemetry should take seriously.
The complaint, brought by 26 Meta employees, seeks a preliminary injunction to stop the company from finalizing separations, an independent audit of the contested tools, and remedies including reinstatement, back pay, lost equity, benefits and other damages. As stated in the filing, the named plaintiffs remain employed until July 22, when their terminations are set to begin unless a court intervenes.
What plaintiffs say
The complaint alleges Meta deployed a “constellation of internal artificial intelligence systems” (plaintiffs’ phrase), including automated “performance” ratings and continuous device monitoring, to score, rank and select employees for inclusion on layoff lists rather than relying on managers who knew the work. It ties several painful examples to that process. A scientist was notified of layoff two days before giving birth, an engineer received a lower rating after time off for an injury, and a manager was dismissed 16 days into medical leave.
“Meta did not assemble the termination list through the considered judgment of managers who knew the work, ” the complaint states. It adds that the systems “in effect penalized the employees for exercising their legal rights to these leaves.”
The complaint describes the monitoring program, reported elsewhere as the Model Capability Initiative (MCI), as capturing keystrokes, mouse activity, browser history, messages, emails and location data from company devices. Plaintiffs say there was no timely opt‑out and that the outputs from multiple internal models were used to create ranked lists of employees.
Meta’s response and internal pushback
Meta disputes the allegation that AI determined layoff lists. A Meta spokesperson said, “These claims lack merit and are not based on facts. Workforce management and organizational decisions were and are made by people, not AI.”
Internal resistance to the monitoring program was already public. More than 1, 600 employees signed a petition opposing the effort (hosted at mcipetition.com), and Mark Zuckerberg announced in June that the program would be paused while the company investigated. As reported by The Information, an internal comment attributed to Zuckerberg explained the rationale behind harvesting employee behavior for models: “The AI models learn from watching really smart people do things. The average intelligence of the people who are at this company is significantly higher than the average set of people that you can get to do tasks.”
How surveillance can translate into disparate harm
“Protected leave” in the complaint refers to time off such as pregnancy/maternity, medical leave and disability accommodations. Telemetry from company devices, like fewer keystrokes, gaps in activity, delayed email replies and location absences, can become model features. If models treat those features as proxies for low productivity and that signal correlates with protected leave, the model can systematically disadvantage employees who exercise lawful leave.
That correlation-versus-causation gap is central to litigation. Plaintiffs must show not only that telemetry correlated with leave, but that the model relied on those telemetry-derived features (or on proxies) in ways that materially affected layoff selections. Defendants can push back by asserting model outputs were advisory, managers retained discretion and any correlations were incidental rather than causal. The court will look for decision trails tying scores to HR actions to resolve that dispute.
What an independent audit should examine, and how
If a court orders an outside review, a useful audit is not a checklist of buzzwords. It should tie technical artifacts to personnel outcomes and explicitly test for disparate impact. Key, actionable areas to inspect:
- Data collection and lineage: ingest policies, device scopes, timestamps, retention windows, and raw telemetry samples (screenshots, keystroke logs, browser history). Confirm whether employees received notice or an opt‑out and which teams or device types were included.
- Model inputs, feature engineering and labels: feature tables, transformation scripts, and training labels. Run correlation matrices and mutual‑information analyses to find proxies for leave. Examine feature importances using explainability tools (SHAP, LIME) to see what drove predictions.
- Output logs and decision trails: archived model scores and rankings with timestamps, and linked HR actions. Look for instances where a score preceded manager action without documented override.
- Access and exposure logs: who queried MCI tables or model artifacts, when, and with what permissions. Verify reports (e.g., Wired as cited by the Guardian) of internal exposure of MCI tables and measure scope of any over‑exposure.
- Statistical disparate‑impact testing: compute selection rates for employees on protected leave versus peers, disparate impact ratios, group calibration checks, and differences in false positive and false negative rates. Run counterfactual simulations to see whether neutralizing leave‑related activity changes rankings.
- Deployment and governance records: CI/CD release notes, model version history, A/B tests, rollout memos, sign‑offs and escalation procedures showing whether humans reviewed or relied on model outputs.
Legal and regulatory context
Several states have tightened oversight of automated decision systems in employment. California, Colorado and Illinois have recently adopted laws or regulations addressing algorithmic decision‑making or automated employment tools. Those frameworks increase the legal exposure companies face when an automated system causes adverse personnel outcomes. Plaintiffs here are asking the court not only to stop separations but to order an independent audit, an outcome that would create a rare, document‑based view into how an AI‑assisted personnel process actually functioned.
What executives and HR leaders should do now
Treat this case as a warning and as a prompt for immediate action. Prioritize three immediate steps and three medium‑term investments:
- Immediate, pause and preserve: suspend use of any automated personnel decision tools until a legal and technical risk assessment is complete. Preserve logs, model artifacts and deployment records for possible audit or litigation.
- Immediate, transparent communication: tell employees what is (and isn’t) collected, how it’s used, and provide opt‑out options where possible. Secrecy amplifies distrust and legal risk.
- Immediate, legal review: involve counsel experienced in employment, privacy and AI/ADSR (automated decision systems) regulation to map exposure and compliance gaps.
- Medium term, governance and testing: institutionalize human‑in‑the‑loop controls, documented override authority and incident response for model errors. Require pre‑deployment bias testing and periodic post‑deployment disparate‑impact audits.
- Medium term, technical safeguards: codify privacy‑minimizing collection, minimize data and aggregate or anonymize where possible. Use explainability tools and counterfactual testing in model validation.
- Medium term, policy and training: update HR policies to prohibit use of protected‑leave signals in scoring unless explicitly accounted for, and train managers to interpret model outputs as advisory not determinative.
Meta’s likely defenses and the evidentiary battleground
Meta is likely to argue that model outputs were advisory, that human managers made final decisions, and that any correlation between leave and telemetry did not drive the outcomes. Plaintiffs will try to produce logs, model outputs, emails, deployment notes or witness testimony showing scores were used operationally or that managers followed ranked lists without meaningful review.
The practical battle will be over artifacts, such as timestamps tying scores to layoff lists, access logs showing HR reliance on model outputs, and internal communications framing automation as decisive. Those are the items an independent audit or discovery would unearth.
Why the case matters beyond Meta
This is not just about one company’s internal tooling. If the complaint’s central claim, that surveillance telemetry plus automated scoring penalized people for taking protected leave, holds up, it will be a concrete example of how product choices and feature engineering can encode legal risk into business processes. Companies that treat employee behavior as neutral training data risk turning lawful leave and disability into de facto negative labels unless governance, transparency and legal safeguards are built in from day one.
Key questions, short, honest answers
-
Did Meta use AI and employee monitoring to select layoffs?
Plaintiffs allege Meta used a “constellation of internal artificial intelligence systems” and continuous device monitoring to score, rank and select employees for layoff; Meta disputes those claims and says workforce decisions were made by people, not AI.
-
What specific data did the monitoring program collect?
According to the complaint and reporting, the program collected keystrokes, mouse activity, browser history, messages, emails and location data from company devices. Reporting about the Model Capability Initiative (MCI) also says some internal tables were exposed to broader internal access, as Wired reported and the Guardian summarized.
-
How many employees brought the lawsuit and when are their terminations effective?
The complaint lists 26 plaintiffs in a 71‑page filing; their terminations are set to begin on July 22 unless the court grants relief, per the filing.
-
What relief are plaintiffs asking for?
They seek a preliminary injunction halting finalization of layoffs, an independent audit of Meta’s tools and processes, and remedies including reinstatement, back pay, lost equity, benefits and damages.
-
What will a credible audit likely reveal?
An audit should surface data‑collection practices, the model features used, feature importances, output logs with timestamps, access records and any manager workflows that used or ignored model outputs, evidence that can show whether automation materially drove personnel outcomes.
The technical problem at the heart of this case is solvable: models and data pipelines can be designed not to encode leave‑related signals as penalties. But solving it requires a mix of engineering, legal review and honest governance, and the humility to stop, preserve evidence and invite independent verification when serious allegations arise.