Meta allegedly left it to its internal artificial intelligence to decide which 8,000 employees to let go, and the criteria reportedly went beyond job performance to include how much each person used the company’s AI tools. That is the picture emerging from a lawsuit filed in California by 26 former employees under pseudonyms, accusing the company of violating federal and state laws by discriminating particularly against those on protected medical or family leave.

At the center of the complaint is an internal architecture Meta is said to have assembled to evaluate staff: a cluster of systems including the ‘Metamate’ assistant, employee-trained ‘second-brain’ agents, keystroke and activity monitoring, token-usage dashboards, and algorithmically assisted performance rankings. According to the lawsuit, employees were graded and grouped into categories such as ‘AI Native,’ ‘AI First,’ or ‘AI Enabled’ based on their adoption of internal tools. In practice, the less you used them, the higher the risk.

The story goes beyond this single legal case. It exposes a systemic risk that comes with deploying AI inside an enterprise, especially when the deployment is on-premise and the models remain beyond any external auditing. In a self-hosted context, the company holds full technical control of the pipeline but also bears near-absolute responsibility for governing the fairness and transparency of the algorithm. If the system functions as a black box—and it often does, because proprietary neural-network rankings are not easily interpretable—opacity becomes a double-edged sword: it safeguards data but can conceal discriminatory automation.

A fresh paradox emerges. An internally managed AI infrastructure is often praised for guaranteeing data sovereignty and regulatory compliance, yet the moment it is used to make decisions that affect people’s rights, that same seclusion becomes a risk factor. No third party validates the criteria, and no established standards certify that an employee-scoring system is fair. If the AI is trained on proxy metrics such as the use of a corporate tool, it can easily penalize those with less digital roles or those legitimately absent, without anyone—except the algorithm—having deliberately decided it.

The lawsuit against Meta thus serves as a wake-up call for anyone designing or managing on-premise deployments. It is not enough to calculate compute power, model latency, or total cost of ownership; auditability mechanisms and human-override capabilities must be embedded from the start when the AI’s output carries legal consequences. The ease with which a company can push an opaque-metric employee evaluation system into production shows how thin the line is between efficient automation and algorithmic drift. And it poses an uncomfortable question: if even the employees of one of the world’s most advanced AI companies become pieces in an incomprehensible ranking, what might happen in businesses now approaching these tools—perhaps with fewer resources to govern them?