AI Bias in Hiring: What Recruiters Need to Know
AI doesn't create bias from nothing — it amplifies and scales bias that already exists in the data it was trained on or the criteria it was given. Used carefully, AI screening can reduce certain human biases. Used carelessly, it can replicate historical discrimination at scale.
How AI Bias Enters Hiring Systems
Three common entry points:
- Training data bias — if an AI screening model was trained on historical hiring data from a company where certain demographics were systematically underrepresented, the model learns to favour profiles that resemble past hires. Historical patterns get encoded as "qualified."
- Rubric bias — if the recruiter defines screening criteria that correlate with protected characteristics (requiring a specific degree from a small set of universities, requiring specific work history that disadvantages career changers), the AI faithfully applies those biased criteria at scale.
- Proxy variable bias — the AI may use variables that aren't themselves protected characteristics but correlate with them. Zip code, employer name, school name, and certain hobbies can serve as proxies for race, class, or gender.
What Reduces AI Bias
- Content-based scoring — evaluating what candidates say in response to questions, not resume patterns, demographic proxies, or presentation style. AI screening that scores interview response content is less bias-prone than AI that scores resumes.
- Rubric review — before deploying a screening rubric, review criteria for disparate impact potential. Requirements that disproportionately screen out qualified candidates from underrepresented groups should be justified by genuine job necessity or removed.
- Regular bias audits — testing screening outcomes by demographic group to identify patterns. If a screening model rejects a demographic group at significantly higher rates, investigating before that becomes a legal issue is essential.
- Human review — AI scores should inform decisions, not make them. Human oversight at the offer stage ensures the final decision can be explained and defended.
The Legal Landscape
Regulation of AI in hiring is accelerating:
- New York City Local Law 144 — requires independent bias audits of automated employment decision tools and public disclosure of results before the tool can be used in NYC hiring.
- Illinois AI Video Interview Act — requires disclosure and consent when AI is used to evaluate video interviews, and prohibits use as the sole basis for employment decisions.
- EU AI Act — classifies AI systems used in employment as high-risk, requiring transparency, human oversight, and documentation of risk management.
- EEOC guidance — the Equal Employment Opportunity Commission has issued guidance stating that employers are liable for discriminatory outcomes from AI tools they use, even if the tool was built by a third party.
Bias Audit vs. General Model Testing
A bias audit examines whether a specific employment selection tool or workflow produces concerning disparities under a defined method, population, time period, and legal context. It is not interchangeable with a generic AI security test, model benchmark, or vendor statement that the system is "unbiased."
Where New York City's automated employment decision tool law applies, employers and employment agencies have specific audit, publication, and notice obligations. Applicability depends on how a tool is used; obtain jurisdiction-specific legal advice rather than presenting a single universal checklist.
Questions to Ask Before Deployment
What specific workflow was tested? Which selection or scoring rates were measured? What data and sample limitations exist? What changed since the audit? Can the hiring organisation examine results and respond to issues?