How Can I Research AI-Assisted Hiring?

What are good research questions about AI in human resource management?

Good research questions name the hiring stage, the system or practice, the group being studied, and the outcome being measured. “Is AI hiring biased?” is too broad to guide a literature review. “How do language-model resume rankings differ by race and gender when qualifications are held constant?” gives the researcher a testable focus.

Build the question from four connected themes. Fairness asks whether scores or selection rates differ across groups. Privacy asks what applicant information is collected, inferred, retained, or shared. Transparency asks whether people know a tool is in use and can understand what it evaluates. Human oversight asks who reviews the output and how that review affects a decision.

Then choose a method that fits the question. An audit can compare outputs across otherwise similar applications. Interviews can explore how recruiters interpret system recommendations. Applicant surveys can examine notice, trust, or perceived ability to challenge an outcome. Mixed methods can connect measured differences with the decisions and experiences surrounding them.

Make the scope explicit: name the hiring stage, population, tool type, comparison, and outcome. For example, a study could ask whether changing names on otherwise comparable resumes changes a language model’s rankings for a defined set of occupations. That question tests identity effects without claiming to describe every employer’s system.

How does AI resume screening create bias in hiring?

AI resume screening can produce different rankings for candidates with comparable qualifications when names, document features, or patterns learned from training data affect model output. An audit can expose those differences, but its findings apply to the systems and conditions tested.

Kyra Wilson and Aylin Caliskan simulated screening across nine occupations using more than 500 public resumes and 500 job descriptions. Their study found that the tested language-model embedding systems favored White-associated names in 85.1% of cases and female-associated names in 11.1% of cases; some analyses found Black male candidates disadvantaged in up to 100% of cases. The authors also identified document length and name frequency in training data as factors affecting selection. These are variables a researcher can test alongside demographic differences.

A separate randomized-resume audit by Jiafu An and colleagues tested several large language models using approximately 361,000 fictitious resumes with randomized social identities. Female candidates often received higher scores, while Black male candidates received lower scores despite comparable qualifications. At a hiring-score threshold, the measured differences could correspond to approximately 1–3 percentage-point differences in hiring probabilities for otherwise similar candidates. The study describes its tested models and experimental setup, not every deployed hiring system.

Bias is not limited to a model’s score. The U.S. Equal Employment Opportunity Commission says federal employment discrimination laws apply to AI used in recruiting, screening, hiring, monitoring, promotion, or termination. It notes that a seemingly neutral practice can raise disparate-impact concerns, including when video software scores applicants poorly because disability affects their speech patterns. A research review should therefore ask what the system measures and whether that measure is relevant to the job.

For a practical framework, compare outcomes across groups, inspect the features that may shape a score, and record the context in which the system is used. NIST identifies fairness, privacy, transparency, accountability, validity, and reliability as trustworthiness concerns. It also stresses that fairness cannot be reduced to demographic balance: the intended use, human judgment, chosen metric, and threshold matter.

What sources can I use for a research paper on AI in HR?

Use peer-reviewed research for empirical findings and public agency sources for legal or measurement context. Start with the two resume-audit studies, then use the EEOC, New York City Department of Consumer and Worker Protection, and NIST to frame legal obligations, transparency, and evaluation methods.

Evaluate each source by asking what was tested, which data were used, what outcome was measured, and what the authors say the results can establish. A controlled audit can isolate identity effects, but it does not prove that every employer’s tool behaves the same way. A public rule can show what covered employers must do in a jurisdiction, but it does not show whether the notice changes candidate experience.

Vendor materials can help identify a product’s stated purpose, features, and claimed safeguards. They are not independent evidence that the tool is fair or effective. Check whether the claim is supported by a published method, accessible data, and outside evaluation. If those details are absent, label it as a vendor claim and avoid treating it as a research finding.

New York City’s Local Law 144 offers a concrete case for studying audit design. For covered tools, the law requires a bias audit within the previous year and specified candidate notices. Its audit includes selection or scoring rates and impact ratios across sex, race/ethnicity, and intersectional categories, with a public summary that describes the data. The rule provides a defined example, not a universal standard for all hiring systems.

Organize a literature review around a comparison rather than a list of papers: what each method reveals, what it misses, and which claims remain unsupported. Audits measure output patterns; interviews can explain how people use those outputs; surveys can capture applicant-reported experience. NIST supports quantitative, qualitative, and mixed approaches, with testing before deployment and during operation, documented methods, and uncertainty. Together, these methods make a stronger research design than a single score or vendor assurance.

For a focused project, state the question, define the evidence needed to answer it, and set a boundary around what the evidence cannot show. A narrow, transparent question is more useful than a broad claim about whether AI hiring is good or bad.