ATS software is broken and everyone knows it

HackerRank open-sourced its applicant tracking system, and a writer immediately stress-tested it by running the same resume multiple times. The result: scores of 90, 74, and 88 on identical input. A 65% failure rate on deterministic scoring. The comments are not kind, calling the grading rubric 'completely stupid' and noting that the system is non-deterministic by design or by accident, which amounts to the same thing for candidates.

The pattern here connects directly to the hiring market signal visible in the YC job listings. Dozens of startups are hiring, but the pipeline for candidates is being filtered by software that scores the same resume differently two-thirds of the time. One commenter noted, dryly, that this is 'actually a fantastic number' compared to human reviewers, which tells you everything about how low the bar is.

The counterpoint: some commenters defend automated screening as a triage tool, not a decision tool. But when the triage is random, the practical effect is that qualified candidates are rejected by coin flip.