In California, twenty-six former Meta employees have filed a federal lawsuit alleging the company used AI-powered systems to select staff for mass layoffs, disproportionately targeting those on medical leave or with disabilities. According to the complaint lodged Monday in Oakland, Meta relied on productivity metrics and AI token usage data to decide who stayed and who went.

It’s not the first time an algorithm has faced legal scrutiny over employment decisions, but the technical detail here is striking: the company allegedly tracked token consumption generated by employees. In the large language model lexicon, tokens are the elementary units of text – words or parts of words – that a model processes. If an organization logs how many tokens a worker produces while interacting with internal tools, that figure can morph into a productivity gauge. Yet such a metric is blind to context: someone slowing down to care for a relative, undergoing treatment, or simply communicating less verbosely risks looking under performing.

The story hits a raw nerve for any enterprise currently evaluating on-premise, self-hosted model deployments. Moving inference and fine-tuning under one’s physical control is often driven by data sovereignty concerns, GDPR compliance, and long-term TCO. But the Meta case shows that technical control isn’t enough – process control is essential. When a company runs an LLM on its own servers to parse internal communications and feed decision-making metrics, it takes full ownership of any discriminatory outcomes. It cannot shift blame onto a cloud provider or a black-box vendor.

The lawsuit signals a broader short circuit. HR departments are adopting AI tools to streamline evaluations, interviews, and even reduction-in-force planning. Yet the transparency of these systems remains low, and their statistical validation is often insufficient. The risk extends beyond legal exposure: there’s a reputational and cultural cost that can erode employee trust. Teams building internal pipelines thus find themselves balancing automation with rights, a tightrope that demands architectures capable of logging, explaining, and correcting automated decisions.

Structurally, the news reinforces a trend: the line between operational efficiency and algorithmic surveillance is thinning. Companies investing in inference hardware – from consumer GPUs to multi-card workstations – should ask not only how many tokens per second they can process, but also what data is being collected, how it’s aggregated, and whether the metrics built atop those tokens can produce distortions. Data sovereignty isn’t merely a shield against external access: it’s also the capability to ensure data remains a management tool and doesn’t turn into a double-edged sword.