Is AI resume screening safe to use?
The technology reads resumes fine. The question is what you let it decide.
Answered in short
5 things that decide this
- 01AI resume screening is safe when it ranks and summarises while a person makes every reject and advance decision.
- 02It becomes legally risky as an automatic rejector, because screening tools fall under employment discrimination law and the employer answers for the tool's disparate impact.
- 03New York City's Local Law 144 already requires independent bias audits and candidate notice for automated employment decision tools, and the EEOC has published guidance applying existing law to algorithmic hiring.
- 04Manual screening is the pressure: recruiters spend 5 to 10 minutes per resume and 80-plus hours per hiring cycle on screening-adjacent work, per figures published by MindStudio. Assistive ranking removes most of those hours without removing the decision-maker.
- 05Safety is auditable behaviour: log what was ranked, on which criteria, and who decided. A staffing firm should be able to explain any candidate's outcome in one query.
The employer owns the tool's mistakes
Discrimination law does not care whether a person or a model produced the disparate impact. If a screening system consistently filters out a protected group, the employer and, for staffing firms, often the agency carries the exposure. Buying the tool does not transfer the liability.
Regulators have started writing this down. New York City's Local Law 144 requires annual independent bias audits and candidate notification where automated tools substantially assist hiring decisions, and other jurisdictions are drafting in the same direction. A firm operating across states should build to the strictest rule it touches.
The practical consequence is architectural: the system must be able to show its criteria and its outcomes by group, which rules out black-box scoring bought on faith.
Rank on stated criteria, decide with people, log everything
Give the system the job's actual requirements and have it rank against those, with a written reason per candidate. Recruiters review the ranked list, and every advance or reject is a person's click. The model compresses reading time; it holds no veto.
Strip what the law ignores and bias feeds on: names, photos, addresses, graduation years. A ranker that never sees them cannot lean on them. Then test outcomes on your own historical data by group before go-live, and quarterly after. Testing is what turns a policy into evidence.
Where the approach is not worth it: low-volume, senior searches where every resume gets read anyway, and roles whose requirements resist written criteria. Assistive screening earns its keep on volume.
Related questions
01Does assistive ranking still trigger bias-audit rules?+
It can, depending on how substantially the tool assists the decision. Local Law 144 turns on whether the tool substantially assists or replaces discretionary decisions. Position the system as decision support, keep humans deciding, and take counsel's view for the jurisdictions you hire in.
02Do candidates have to be told?+
In New York City, yes, with advance notice. Elsewhere the rules vary and are moving. Disclosure costs little and reads as fairness, so the practical answer is to notify everywhere and argue nowhere.
03Is a custom-built screener safer than an off-the-shelf one?+
It is more inspectable, which is the property that matters: your criteria, your logs, your outcome data, testable on demand. Off-the-shelf tools can be compliant too, if they expose the same evidence. The unsafe option in either case is the one whose reasoning nobody can see.
