AI Interview Bias and Fairness (2026): How Hiring Teams Make AI Interviews Fair and Compliant

AI interviews can be biased, and so can human interviews. The question for a hiring team is not whether bias is possible but whether you can measure it, limit it and show your work. The practical controls are the same ones that make any interview fairer: a written rubric, independent bias audits, clear candidate notice, an accommodations path, and a human who reviews and owns the decision. This guide covers how bias arises in AI interviews, what each control does, and where the main rules stand as of October 2026. It is also written for candidates who searched "ai interview bias" and want to know what protects them.

This is general information, not legal advice.

At a glance

Question	Short answer
Can AI interviews be biased?	Yes. Scoring models, training data, rubrics, speech recognition and question design can all create unequal outcomes.
Are they more or less biased than human interviews?	It depends on the tool and the process. Consistency can help; poorly built or unaudited tools can scale bias. Evidence is tool-specific, so audit yours.
What is the most effective safeguard?	A structured rubric scored on job-related dimensions, combined with human review and ongoing monitoring of outcomes.
What does NYC Local Law 144 require?	A bias audit within the year before use, a public summary and candidate notice for covered automated employment decision tools.
What is the EEOC's position in 2026?	It removed its AI guidance documents in January 2025 and stopped pursuing disparate-impact investigations in late 2025, but the underlying laws still apply.
What does the EU AI Act say?	Recruitment and candidate-evaluation AI is high-risk; the applicable date for those obligations was delayed to 2 December 2027.
Do candidates have rights?	Depending on location: notice, consent, an alternative process, accommodations and data rights. Ask the employer.

Where bias in AI interviews comes from

Training data and scoring models. If a model learned from past hiring decisions, it can reproduce past patterns. Models trained on narrow groups of speakers can perform worse for others.

Speech and language handling. Voice-led interviews depend on transcription. Accents, speech differences, stammers, background noise or non-native phrasing can degrade transcripts, which can degrade scores if the scoring reads the transcript. Ask vendors how this has been tested.

Proxy features. Even if a tool does not use protected characteristics, it can use proxies, such as location, school or vocal characteristics that correlate with them. The EU AI Act already prohibits inferring emotions of people in the workplace, with limited medical and safety exceptions, so ask whether a tool infers emotion or personality from voice or face at all.

Rubric design. If the rubric rewards style (fluency, confidence, speed) over substance, it may disadvantage people who communicate differently without being less capable.

Question design. Questions that assume particular cultural context, or that elicit health or disability information, can create unequal exposure. California's regulations, per law-firm summaries, note that some automated assessments that elicit disability information can amount to unlawful medical inquiries.

Process design. Auto-rejecting on a score without human review turns any model error into a hiring decision.

The controls that actually help

1. Structured rubrics

A structured interview asks every candidate the same core questions and scores answers against written criteria for each dimension. This is the oldest and best-supported fairness practice in hiring, and it applies to AI interviews as much as human ones. In practice:

• Define dimensions tied to the job (for engineers: problem solving, code correctness, communication, design trade-offs).
• Write what a strong, borderline and weak answer looks like for each dimension.
• Require that scores cite evidence from the interview, not impressions.
• Exclude things that are not job-related, such as accent, appearance or filler words.

HireVue's AI Interviewer page, for example, describes scores tied to behavioral anchors; ThitaHire's site describes dimension-wise scoring with strengths and risk markers and a consistent report structure. Whatever vendor you use, ask to see the rubric and test it. See our buyer's guide for how to test rubric consistency.

2. Bias audits and ongoing monitoring

A bias audit compares outcomes across groups. Under NYC Local Law 144, the audit calculates selection rates and impact ratios for sex and race/ethnicity categories, per law-firm summaries of the final rules. Practical guidance:

• Ask the vendor for the latest independent audit: auditor, date, data used, categories tested and the exact tool version.
• Do not assume a vendor's audit satisfies your own obligations. Under the NYC law the employer must ensure a current audit exists and is published.
• Monitor your own pass-through rates by group where you can lawfully collect that data, and have someone qualified interpret them. A common screening heuristic is the four-fifths rule from US federal selection guidelines, but it is a flag for investigation, not a legal safe harbor.
• Re-audit when the model, rubric or candidate population changes.

Audits have limits. In December 2025 the New York State Comptroller found that the city's enforcement of Local Law 144 had been ineffective, with problems in complaint handling and compliance review. That suggests compliance may be checked more closely over time, and that employers should not rely on weak enforcement.

3. Candidate notice and consent

Tell candidates before the interview that AI is used, what it evaluates and what happens to their data. Requirements vary:

• NYC: Candidates must receive notice before a covered tool is used. The city's slides, per its AEDT page, say 10 business days. Law-firm summaries add that candidates may request an alternative process or accommodation and that data practices must be disclosed.
• Illinois: The AI Video Interview Act requires notice, an explanation of how the AI works and consent before AI analyzes video interviews. Amendments to the Illinois Human Rights Act effective 1 January 2026 require notice when AI is used in covered employment decisions.
• EU: The AI Act's transparency duties and, for high-risk systems, obligations to inform affected people.

Good notice is short, specific and written in plain language. Include who to contact with questions.

4. Accommodations and alternatives

Candidates with disabilities, speech differences or connectivity problems may be disadvantaged by a voice-led, timed or camera-based format. Build a visible path to request:

• extra time or breaks;
• a typed or human-led alternative;
• a different time or device;
• help with technical failure.

Decide in advance who handles requests and how fast. Do not penalize candidates who use the path. The ADA still requires reasonable accommodations, even though the EEOC removed its AI-specific guidance on the topic in 2025.

5. Human review of decisions

AI output should inform a decision, not make it. Name the person who reviews each AI report, what they see (evidence, not just a score), and what they may override. Review borderline and flagged cases first. Keep records of overrides; they are a free source of signal about where the tool and your standards disagree. HireVue's AI Interviewer page frames scores as inputs to recruiter judgment; the same principle should hold for any tool you use.

6. Integrity without over-policing

Integrity measures can create their own bias: flagging behavior (looking away, long pauses, background noise) may disproportionately affect candidates with disabilities, neurodivergent candidates or those in shared homes. Escalate flags to a person, never auto-reject on them, and tell candidates what is monitored. Our guide on whether online assessments are proctored explains typical monitoring from the candidate side.

What the rules say in October 2026

NYC Local Law 144

Covered employers and employment agencies must not use an automated employment decision tool to evaluate NYC candidates unless a bias audit was done within the preceding year, results are publicly summarized and candidates receive notice. The Department of Consumer and Worker Protection enforces it, with enforcement beginning 5 July 2023. Law-firm summaries (Perkins Coie) describe the audit methodology and exclusions. Check the city's rules for the exact definition of a covered tool and notice contents.

EEOC guidance status

According to law-firm reporting, the EEOC removed its AI technical-assistance documents (including its 2023 Title VII and 2022 ADA documents) from its website on 27 January 2025, and on 15 September 2025 directed field offices to administratively close disparate-impact investigations by 30 September 2025. In the iTutorGroup age-discrimination case, the agency withdrew a disparate-impact allegation about an AI screening tool. The statutes themselves (Title VII, the ADA and the ADEA) are unchanged, private plaintiffs can still sue, and state laws are unaffected. The case of Mobley v. Workday, in which a federal court granted preliminary collective certification on an age-discrimination claim over AI screening in May 2025, was still pending in reporting from spring 2026. In short: less federal guidance does not mean less legal exposure.

EU AI Act

Annex III classifies AI systems intended for recruitment or selection of people, including evaluating candidates, as high-risk. Providers and deployers of high-risk systems face requirements including risk management, human oversight, logging and transparency. The Digital Omnibus on AI (Regulation (EU) 2026/1744, published 24 July 2026) moved the date for standalone Annex III systems from 2 August 2026 to 2 December 2027, per law-firm summaries. It is a delay, not a repeal, so employers hiring in the EU should still prepare. Verify dates in the Official Journal.

California, Illinois and Colorado

California's Civil Rights Council regulations on automated-decision systems took effect 1 October 2025, and per law-firm alerts require retaining related records for at least four years. Illinois's amendments took effect 1 January 2026, with implementing rules still being developed in reporting we found. Colorado's AI Act has been delayed more than once, and we could not confirm its current effective date, so check directly.

A fairness checklist for hiring teams

1. Write the rubric and make sure every dimension is job-related.
2. Get the vendor's latest independent bias audit and confirm it covers your use case.
3. Map where your candidates live and which laws apply.
4. Draft candidate notice and consent language and a plain-language explanation of what is evaluated.
5. Publish an accommodations and alternative-process path.
6. Assign human reviewers and define what they can override.
7. Never auto-reject on an integrity flag.
8. Monitor outcomes by group and review them on a schedule.
9. Keep records, including notice, audit documents, overrides and complaints, for the period your laws require.
10. Take the interview yourself and ask someone with a different accent or setup to do the same.

If you are a candidate

You can ask the employer: Is AI used in this interview? What does it evaluate? Is there an alternative or an accommodation process? Who reviews the result? How long is my data kept? You are entitled to ask, and in some places to receive, answers. Practicing in the same format helps you perform at your best; see how to pass an AI interview and try Thita's free AI mock interview. For which companies use AI interviewers, see companies using AI interviewers in 2026.

Where ThitaHire fits

ThitaHire is our AI interview platform for hiring teams. Its site describes role-specific rounds, voice-led adaptive interviews, dimension-wise scoring, structured reports with strengths, concerns and a recommendation signal, and hiring analytics. The site does not name compliance certifications, bias-audit documents, accommodation workflows or proctoring features, so if fairness and compliance are your main concern, ask the team for those materials and for how notice, alternatives and human review would work in your process, the same as you would with any vendor.

FAQ

Is AI interview bias real? Yes. Models, speech recognition and rubrics can create unequal outcomes. The risk is manageable with audits, structured rubrics and human review, but it does not disappear.

Are AI interviews fair? They can be fairer than unstructured human interviews when they apply the same rubric to everyone and are audited; they can be less fair if poorly built or left unmonitored. Fairness is a property of a specific tool and process.

Is AI interviewing legal? Generally yes, with conditions that vary by location, such as notice, bias audits or consent. NYC, Illinois, California and the EU all have rules. Confirm with counsel.

Do AI interviews discriminate against accents? It is a known risk for voice-based tools. Ask the vendor how transcription and scoring were tested across accents and speech patterns, and provide a typed or human alternative.

Can I ask for a human interview instead of an AI interview? You can ask. Whether the employer must provide one depends on location and the tool; in NYC, law-firm summaries say candidates may request an alternative process or accommodation. Ask the employer.

What is a bias audit for AI hiring? An independent analysis comparing how a tool's selection or scoring outcomes differ across groups. In NYC it reports selection rates and impact ratios by sex and race/ethnicity categories.

Does the EEOC regulate AI hiring tools in 2026? The EEOC removed its AI guidance in 2025 and has largely stopped disparate-impact investigations, but federal anti-discrimination laws still apply and private lawsuits continue. Check current guidance with counsel.

How do I reduce bias in AI interviews? Use a job-related rubric, audit the tool, give candidates notice and alternatives, keep a human in the decision, and monitor outcomes regularly.

See the evidence behind the score

If you are evaluating AI interviews with fairness in mind, ask to see what a hiring manager would see for a real candidate: the rubric, the evidence and the recommendation. Try the ThitaHire live demo or contact the team to walk through your fairness and review requirements.