AI-Assisted Resume Screening with Bias Audit

Problem class

Manual screening is biased and slow; pure AI screening (Amazon's failed tool) compounds historical bias. The hybrid model — AI-assisted with rigorous bias governance — speeds screening while reducing both bias and discrimination liability.

Mechanism

AI ranks candidates against role requirements using gamified assessments and structured interviews. Bias audits run continuously with the four-fifths rule across protected categories. Audit-ai libraries quantify disparate impact. Humans make all final hiring decisions; AI provides shortlists, not decisions. NYC LL 144 audit and disclosure compliance is mandatory.

Required inputs

• Diverse training data with bias remediation
• Job competency definitions
• Bias audit infrastructure (audit-ai or equivalent)
• Candidate consent and disclosure workflow
• Human reviewer pool for final decisions

Produced outputs

• Ranked candidate shortlists
• Continuous bias audit results
• Compliance documentation (NYC LL 144, EU AI Act, Colorado)
• Time-to-hire reduction

Industries where this is standard

• Large public companies in NYC (LL 144 jurisdiction)
• EU employers (EU AI Act high-risk classification, August 2026)
• Federal contractors under OFCCP
• High-volume hiring environments (retail, hospitality, gig)
• Companies with global graduate recruiting programs

Counterexamples

• Deploying AI screening without bias audit — Amazon scrapped its 2014-2018 tool after discovering it penalized "women's" resumes and couldn't verify bias removal. The tool itself was the liability.
• Facial analysis components — HireVue removed its facial analysis after EPIC FTC complaint, finding it contributed only 0.25% to predictive power; the regulatory and reputational risk vastly exceeds the marginal value.

Representative implementations

• Unilever — Pymetrics (now Harver) + HireVue; 90% reduction in time-to-hire (4+ months → ~4 weeks), £1M annual savings, 50,000 hours saved, 16% diversity increase. Process: gamified assessments → AI-scored video interviews → human-led assessment centers.
• Pymetrics — first-of-its-kind third-party bias audit by Northeastern University (FAccT 2021); met the EEOC four-fifths rule across gender and race; open-sourced audit-ai Python library; completed NYC LL144 audit June 2023.
• NFL — Greenhouse AI-powered filtering with structured fallback; 24% time-to-fill reduction (63 → 48 days) within structured interview framework.

Common tooling categories

AI ranking engine + bias audit framework + structured assessment platform + audit logging + human review workflow.