It is true, in the narrow sense in which these things are usually true, that the algorithm has no intent. An internal performance-ranking system that relies on keystroke-activity data, AI-token-usage dashboards, and continuously-updated algorithmic scores does not know that an employee is on medical leave, parental leave, or disability accommodation. It knows only that the data stream has stopped. The system is not malicious. It is structurally blind. The trouble is that the law requires a human to notice the blindness and pause the machine.
The 26 employees who filed suit in Oakland this week are not claiming that Meta’s AI singled them out by name. They are claiming that Meta failed to do what the law requires: pause the automated process for the individualized, leave-neutral review that federal and state protections mandate. The system, by design, cannot accumulate scores for an employee who is not producing keystrokes and tokens. The company knew this, or should have known it, and did not stop the machine.
This is the shitty technology adoption curve — Doctorow’s term for the pattern — in a clean, well-lit office. Bossware, the continuous computer-mediated surveillance of every keystroke, every API call, every token generated, climbs the privilege gradient. It was tested on prison labor, then warehouse workers, then customer-service representatives working from home. Now it has reached the desks of software engineers in Menlo Park, and the mechanism is the same: if you stop producing data, the system marks you as dead weight.
Meta’s May announcement of 8,000 layoffs was framed as an efficiency drive; the company has been explicit about AI as a headcount-reduction tool. The efficiency drive is real. What is in question is whose efficiency, and at whose cost. An algorithm that selects workers for termination is efficient the way a guillotine is efficient: it processes the queue without regard to which necks are under the blade.
The legal question is whether Meta violated federal and state leave protections and disability-accommodation law. The engineering question is more fundamental: can an employer build a performance evaluation system that is structurally incapable of accounting for leave, and then claim the system’s output is performance-neutral? The answer is no, and the engineers who built it knew it. The problem is not a bug in the algorithm. The problem is that the algorithm was designed to measure what was measurable — continuous activity — and the company chose to treat the measurable as the complete picture.
Doctorow calls this the reverse-centaur: the human pressed into service as a peripheral for the machine, running at the machine’s pace. The Meta employees on leave are not running at any pace. They are recovering from surgery, caring for a newborn, managing a disability. The machine cannot see them. The company chose not to look.
The straightforward remedy is a regulatory requirement that automated employment-termination systems be subject to human review for any employee on protected leave, and that the systems themselves be designed to flag leave periods for exclusion from performance scores. The Federal Trade Commission has signaled it may use its enforcement powers to address algorithmic discrimination in employment — a rulemaking could set clear standards. It should. Deadlines are the only part of regulatory processes that the regulated actually respect, and the clock hasn’t started.
The workers’ separations are scheduled to begin July 22. The lawsuit will take months, years. The algorithm, in the meantime, is still running. The next round of scores will be computed, the next list of names will be generated, and the system will still be unable to distinguish between a worker who is not productive and a worker who is not here. In the absence of a deadline, the machine decides.