Algorithmic advice on how to get approved next time has a shelf life, and new research shows it can expire the moment other rejected applicants act on it too.
Researchers studied algorithmic recourse, the practice of telling someone denied a loan, a job, or any other automated decision exactly what minimum changes would flip the outcome. That advice assumes the model's acceptance bar stays fixed while people act on it, but when applicants are competing for a limited number of approvals and many act at once, the cutoff score itself can move. The shift can invalidate the original advice, since hitting the old target no longer guarantees acceptance. The researchers built a framework called recourse under competition that uses the Implicit Function Theorem to set recommendations and score targets accounting for the shift, trading recourse cost against the odds the advice still works once the bar moves.
This matters anywhere recourse advice meets scarcity, such as lending, hiring pipelines, or school admissions. A system that ignores crowd effects hands out instructions that look sound for one person and backfire for the group. Tests on synthetic and real datasets found personalizing the score target per applicant raises success rates but costs more effort, while a single shared target is cheaper and holds up reasonably well at low to medium success rates.
It's a useful reminder that explainable-AI advice is also a move in a competitive game, and the rules shift the moment everyone plays the same hand.