AI for Recruiters and HR Teams: Where Tools Help and What to Verify

AI helps recruiters most with the work around a hiring decision: finding candidates, scheduling, capturing interview notes and answering routine questions.

It carries the most risk when it ranks or rejects people, because that is where bias and legal exposure concentrate.

Two bodies of research frame the choice. One shows how language models can inherit bias when they rank resumes, and the other shows an AI interviewer improving hiring outcomes in a large field experiment.

This guide covers both, maps AI to each stage of hiring, walks through the adverse-impact check with a worked example, summarizes the rules that apply, and lists what to ask any vendor.

What the Research Says: One Warning and One Promising Result

In 2024, University of Washington researchers tested how three language models ranked resumes. They took more than 550 real resumes, swapped in 120 first names associated with white and Black men and women, and ran more than three million comparisons against job descriptions.

The models favored resumes with white-associated names in 85.1% of cases and female-associated names in 11.1%.

Resumes with Black male names were never favored over identical resumes with white male names, according to the paper and the university's summary.

The study has limits. It tested open-source embedding models in a research setting, and it changed only names, so it shows what a model can do and says little about any particular commercial product.

One of the authors noted that removing names would not solve the problem, because models can infer identity from schools, locations and word choices.

A different study points the other way. In a field experiment published as Voice AI in Firms, Brian Jabarian and Luca Henkel randomly assigned about 70,000 applications at a recruitment outsourcing firm to an AI voice interviewer, a human interviewer or a choice between the two.

Applicants interviewed by the AI were 12% more likely to receive an offer, 18% more likely to start and 17% more likely to stay 30 days.

When given the choice, 78% of applicants picked the AI interviewer. Human recruiters still reviewed the results and made the hiring decisions.

The caveats matter here too. The study covered one firm and mostly high-volume entry-level roles, and the lead author holds an unpaid advisory role at the firm.

The authors say it is uncertain whether the results carry over to other countries or higher-skill jobs. The absolute rates were also small, so the percentages describe changes from a low starting point.

Read together, the studies suggest that the design and the job matter. A model that ranks people can reproduce bias, and a tool that standardizes how information is collected, with a person deciding, may help.

Also Read: The Ethics and Risks of AI in the Workplace: What Every Business Needs to Know

Where AI Fits at Each Stage of Hiring

Vendor guides sort recruiting tools into a few categories. The tools below are the ones those guides name most often, and the order follows the hiring process.

• Sourcing: SeekOut, hireEZ, Findem and Gem use AI to search and rank profiles. Ask what the ranking is based on, since the first study shows how rankings can inherit bias.
• Job posts and outreach: Assistants can draft job descriptions and first messages. Review them for exclusionary wording and for claims about pay, flexibility or requirements that must be accurate.
• Screening and matching: Applicant tracking systems such as Greenhouse, Ashby and Lever, and talent-intelligence tools such as Eightfold, score or rank applicants. This is the highest-risk stage, so keep a person responsible for the final call.
• Scheduling and candidate chat: Paradox and Humanly handle scheduling, questions and early screening conversations by chat or text. Scheduling carries little risk, while screening questions must relate to the job and work for candidates with disabilities.
• Interview notes: Metaview and BrightHire record interviews, summarize them and map them to scorecards. Consent rules apply to recordings, and summaries need checking.
• Onboarding and employee questions: HR assistants answer policy and benefits questions. A wrong answer about leave or pay creates real problems, so link them to the source policy.

Pricing for these tools is mostly quote-based and sources conflict, so request current figures from each vendor. Our guide to AI note-taking and AI research tools, linked below, cover recording consent and checking AI summaries.

Also Read: AI Note-Taking for Meetings and Lectures: How to Choose and What to Check

The Adverse Impact Check, Worked Through

The standard first test for a hiring tool is simple arithmetic. It compares how often each group of applicants passes a stage, and it applies to a resume screen, a skills test or an AI ranking alike.

Under the federal Uniform Guidelines, a selection rate for any group that is less than four-fifths, or 80%, of the rate for the group with the highest rate is generally regarded as evidence of adverse impact. New York City's audit rules use a similar impact ratio. The example below is hypothetical:

• Group A: 200 applicants and 120 advanced, a selection rate of 60%.
• Group B: 100 applicants and 36 advanced, a selection rate of 36%.
• Impact ratio: 36% divided by 60% is 0.60, below the 0.80 threshold.

A ratio of 0.60 raises a flag rather than settling the question, and it means the stage needs a closer look. The next questions are whether the criteria relate to the job, whether a less discriminatory alternative exists, and whether the gap holds up statistically.

Three cautions apply. The four-fifths rule is a rule of thumb that carries no legal safe harbor, so a tool can pass it and still raise legal problems.

Small groups produce unstable ratios, because a single hire moves the number a lot. And the check works only if you record who applied and who advanced at each stage.

Run it at every stage that uses AI, not only at the final hire, and rerun it regularly.

The Rules That Apply

Rules in this area are changing quickly, and they differ by location. As of fall 2026, these are the main ones named in law-firm and compliance summaries.

• US federal: Title VII, the ADA and the ADEA apply to AI-assisted decisions as they do to any others. The EEOC removed its AI guidance from its website on January 27, 2025, though the underlying laws did not change.
• New York City: Local Law 144 requires a bias audit within a year before an automated tool is used, a public summary of the results and advance notice to candidates. A December 2025 state audit found significant enforcement gaps.
• California: Civil Rights Council regulations in force since October 1, 2025 apply anti-discrimination law to automated decision systems and require records to be kept for four years.
• Illinois: Amendments to the Human Rights Act, in force since January 1, 2026, bar AI that has a discriminatory effect and require notice to applicants and employees.
• Colorado: Its AI law has been replaced, and one 2026 summary reports a new start date of January 1, 2027.
• European Union: Hiring tools count as high-risk under the AI Act, and the main obligations have been postponed to December 2027, as Gibson Dunn's summary describes.

Sources for the rest include a Rework overview of AI hiring laws and a June 2026 National Law Review briefing. Confirm current requirements with employment counsel before you rely on any of them.

Courts are also starting to examine vendors directly. Our ethics guide, linked above, covers the lawsuit in which a federal court let claims proceed against an AI vendor itself.

Also Read: How AI Chatbots Are Revolutionizing Customer Service

AI-Written Applications and Fake Candidates

Recruiters now face a second AI problem, which comes from the candidate side. A Gartner survey of 3,290 candidates in late 2024 found that 39% used AI during the application process, mostly to write resume text. That is a normal development, and it makes polished applications less informative.

The more serious issue is fraud. In a Gartner survey of 3,000 candidates, 6% admitted to interview fraud, either posing as someone else or having someone pose as them.

Gartner predicts that by 2028, one in four candidate profiles worldwide will be fake, according to its press release and HR Dive's coverage. That is a forecast and no measurement, so treat it as a reason to plan.

Gartner recommends a layered response, and these steps fit its advice.

• Set the rules: State what AI use is acceptable in applications and assessments, so candidates know where the line is.
• Verify identity: Confirm identity at a later stage through a document check or an in-person step.
• Check history at the source: Contact employers and schools using numbers you find yourself, rather than the ones the applicant supplied.
• Add a live task: Include a short, unscripted, job-related exercise on camera or in person.
• Train reviewers: Make sure interviewers know what to look for and where to escalate a concern.

Apply these steps to every candidate, and provide accommodations on request, because extra checks that fall unevenly on some groups can create their own fairness problem.

Resist the shortcut of rejecting applicants because a detector flagged their writing as AI-generated. Detectors make mistakes, including false positives on real writers, and our detector comparison, linked below, covers the evidence.

Also Read: 7 AI Content Detectors Compared: What They Catch and Miss

Questions to Put to Any Vendor

A demo shows what a tool does when it works. These requests show how the vendor handles the risks above, and the answers are worth having in writing.

1. Bias audit results: Ask for the most recent independent audit, including which groups it covered and the impact ratios.
2. Validation: Ask how the vendor shows the tool predicts performance in jobs like yours.
3. Inputs: Ask which data the tool uses, and whether names, photos, postal codes or schools can influence its output.
4. Explanations and overrides: Ask how a reviewer can see why a candidate was ranked and change the outcome.
5. Notices and alternatives: Ask what candidate notices the tool supports and how candidates can request another process or an accommodation.
6. Data handling: Ask where recordings and resumes are stored, for how long and whether they are used to train models.
7. Record keeping: Ask how the vendor will help you keep the records that audits and regulators may require.

Monitor After Launch, Not Only Before

A tool that passed an audit last year can behave differently on this year's applicants, because job requirements, candidate pools and the models themselves all change.

• Selection rates by group at each stage: Rerun the adverse impact check on a schedule, such as quarterly.
• Override rate: Track how often reviewers disagree with the tool, and read a sample of the disagreements.
• Candidate complaints: Treat every complaint about the process as data, especially from candidates who asked for an accommodation.
• Outcome checks: Compare the performance and retention of people hired through the tool with those hired another way.
• Vendor changes: Ask to be told when the vendor changes a model or scoring method, and treat a major change as a reason to retest.

Keep a person accountable for each of these checks, and write down what you find.

Decisions to Keep in Human Hands

Some decisions carry enough weight that a person should make them and be able to explain them.

• Final hire and reject decisions: A person who can explain the reasoning should own each outcome.
• Accommodation requests: Judgments about disability and reasonable adjustments need human judgment and care.
• Removing a flagged candidate: If an AI tool flags a candidate for fraud or misconduct, a person should review the evidence before acting.
• Explanations to candidates: When someone asks why they were not selected, a person should be able to answer.

Tell Candidates What You Are Doing

Candidates deserve to know when AI is part of the process, and several rules now require it. Gartner found that only 26% of candidates trust AI to evaluate them fairly, so openness also protects your ability to attract people.

A short notice that names the tool's role, such as scheduling, interview notes or scoring, and explains how to ask for a different process, meets the spirit of most of them.

The field experiment offers a useful signal. When applicants could choose between an AI and a human interviewer, 78% chose the AI, and applicants with lower test scores were more likely to choose it.

Offering a choice respects preferences, and a human alternative also matters for candidates with disabilities who may need one.

Interview recordings need consent from the people recorded, and our note-taking guide covers how to ask. Say how long recordings are kept and who can see them.

Also Read: AI Tools for Resume Writing and Job Hunting: Matched to the Task

Conclusion

AI can reduce the routine work in recruiting and improve the consistency of interviews, and it can reproduce bias when it ranks people. The difference lies in where you use it, how you check it and who makes the final decision.

Use AI most freely for scheduling, notes and candidate questions, test every stage that screens or ranks against the adverse impact check, ask vendors for evidence in writing, tell candidates what you are doing, and keep a person accountable for each hiring decision.

Also Read: AI Research Tools: How to Find, Summarize and Verify Information