The Accrual Trap: When AI Workplace Scoring Picks the People Leave Was Meant to Spare

AI workplace scoring is being asked to pick who gets fired, and the workers it picks are the ones protected leave was designed to spare. Productivity systems run on a hidden premise: that the score is a fair picture of the worker. It is not — it is a picture of the work the worker was present to perform. Protected leave breaks that premise by design, and the systems shipping in 2026 do not adjust for it.

Call it the accrual trap: AI scores the very output a leave-taker is shielded from producing, then exports the lowest scorers into the termination queue. The Business Insider report on Meta's federal complaint names the mechanism. Twenty-six plaintiffs allege Meta's internal systems — Metamate, AI-usage dashboards, activity-monitoring software, an 'AI Native / AI First / AI Enabled' ladder — measured work output, software-development activity, and AI-tool usage that a worker on medical, parental, or disability leave cannot accrue. The complaint's line is the load-bearing one: the termination list, the filing says, was not assembled 'through the considered judgment of managers who knew the work.'

The mechanism generalizes. A firm deploys accrual-based scoring. A worker takes leave and accrual halts. The leave-taker's score falls. HR pulls the bottom of the distribution. The system prints a defensible paper trail. The list assembles itself; malice is optional.

The common read is that AI 'targeted' leave-takers. The pattern underneath is quieter and worse: the system did not aim at anyone. It measured exactly what leave was designed to protect you from producing, and selection fell out as math. 'Human-in-the-loop' only survives if the loop happens before the model builds the list, not after. The accrual trap is the new default.