The most valuable use of machine learning in screening is not making decisions, it is matching identities. Court and county records are notoriously messy: inconsistent name formats, missing middle initials, transposed dates of birth. Models that resolve whether two records describe the same person, and flag near-matches for human review, measurably reduce both false positives and missed records.
AI is also good at classification and extraction: reading an unstructured disposition line and identifying whether a charge was a felony or misdemeanor, dismissed or convicted. This speeds up review, but it should surface the source text alongside the label so a human can verify it.
What AI should not do is decide who gets hired or housed. An opaque score that rejects applicants is a compliance liability, it is hard to explain in an adverse-action notice and nearly impossible to defend against a disparate-impact challenge. Regulators have signaled that "the algorithm did it" is not a defense.
The honest framing is this: AI makes screening faster and more accurate at finding and organizing records. The adjudication, the judgment about relevance to a specific role, belongs to a documented, human-governed policy.