TL;DR: CV screening is one of the most bias-prone steps in hiring. Landmark research found that identical resumes with different names received callback rates that differed by roughly 50%. AI-scored async video interviews reduce that bias by giving every candidate identical questions, identical conditions, and one consistent scoring rubric, while a human recruiter still makes every decision. To deploy AI fairly, validate job-relatedness, test for adverse impact, avoid emotion recognition (prohibited in the workplace under the EU AI Act), and keep human oversight to satisfy GDPR Article 22. Done right, AI makes hiring measurably fairer than the status quo, not less.
Why Is Traditional CV Screening So Biased?
Before asking whether AI is fair, it is worth being honest about the baseline it replaces. Manual CV screening fails almost every fairness test we have.
The most cited evidence comes from Bertrand and Mullainathan, who sent thousands of identical resumes to real job openings, changing only the applicant's name. Resumes with white-sounding names received about 50% more callbacks than identical resumes with names perceived as Black. The skills were the same. The outcome was not.
It gets worse when you look at how CVs are actually read. Eye-tracking research from Ladders found recruiters spend an average of 7.4 seconds on an initial resume scan. In 7.4 seconds, nobody evaluates competence. They pattern-match on proxies: university names, employer brands, employment gaps, postcodes, even hobbies. Every one of those proxies correlates with socioeconomic background, gender, age, or ethnicity.
Add inconsistency (two recruiters ranking the same CV pile rarely agree) and you have a screening method that is fast, familiar, and structurally unfair. That is the honest benchmark any AI system should be measured against.
How Do AI-Scored Interviews Reduce Bias Compared to CV Screening?
Async video interviews with AI scoring attack bias through three concrete mechanisms.
1. Every candidate gets the exact same structured interview
In an async interview, the hiring team defines the questions once. Every candidate answers the same questions, in the same order, with the same preparation and response time limits. This is structured interviewing by construction, and the evidence behind structure is among the strongest in personnel psychology: Schmidt and Hunter showed structured interviews are roughly twice as predictive of job performance as unstructured ones, and meta-analytic work by Aamodt et al. found unstructured interviews are about 2.5x more biased. Live interviews drift toward small talk and improvised questions. Async interviews cannot drift.
2. Scoring evaluates what candidates say, not who they resemble
AI scoring in a well-designed platform works from a job-relevant rubric: does the answer demonstrate the competency the question targets? The model applies the same rubric to candidate 1 and candidate 200 with identical patience. It does not get tired at 4 pm, does not favor candidates who share the reviewer's alma mater, and does not shortlist based on a name or a photo. Crucially, fair AI scoring evaluates the content of answers. It should never infer emotions or personality from facial expressions, a practice the EU has now explicitly banned in employment contexts (more below).
3. Humans decide, with better evidence in front of them
AI scores rank and summarize. People decide. In StormInterview's swipe-review workflow, a recruiter reviews AI-scored responses side by side and swipes to advance or reject each candidate. That human-in-the-loop design matters for two reasons. First, it is what fairness requires: an algorithm should surface evidence, not issue verdicts. Second, it is what European law requires, because GDPR restricts solely automated decisions with significant effects on individuals. The recruiter stays accountable for every outcome, but works from three minutes of structured, job-relevant evidence per candidate instead of a 7.4-second CV glance.
What Should You Validate Before Trusting an AI Score?
"We use AI" is not a fairness claim. Validation is. Before relying on any AI-scored assessment, verify five things:
- Job-relatedness: Every question and scoring criterion should map to a competency the role actually requires. If you cannot explain why a criterion predicts performance in this job, remove it.
- Adverse impact testing: Compare pass-through rates across demographic groups. A common benchmark is the four-fifths rule: if any group's selection rate falls below 80% of the highest group's rate, investigate. New York City's Local Law 144 already mandates independent bias audits for automated employment decision tools, and EU deployers should expect similar scrutiny.
- Explainability: A recruiter should be able to see why a response scored the way it did, tied to the rubric, and override it. Black-box scores that cannot be challenged cannot be audited.
- No facial analysis or emotion inference: Scoring should rest on what candidates say, not micro-expressions, tone-of-voice personality claims, or other pseudo-signals with weak scientific support.
- Ongoing monitoring: Bias is not a launch-day checkbox. Track score distributions, override rates, and outcomes over time, and re-test whenever the model, questions, or candidate pool changes materially.
What Does the EU AI Act Mean for AI in Hiring?
The EU AI Act (Regulation 2024/1689) classifies AI systems used for recruitment and selection, including systems that filter applications or evaluate candidates in interviews, as high-risk under Annex III. High-risk does not mean prohibited. It means regulated, with obligations that broadly apply from 2 August 2026.
If you deploy AI-scored interviews in the EU, the practical obligations include:
- Human oversight: Trained, competent people must supervise the system and be able to override its outputs. A swipe-review step where recruiters confirm or reject every advance is exactly this control.
- Transparency to candidates: Applicants must be informed that an AI system is used in their evaluation.
- Monitoring and logging: Deployers must monitor the system's operation and retain logs, which also gives you the audit trail you need for bias testing.
- Using the system as intended: Follow the provider's instructions for use and feed it relevant, representative input data.
One rule is already in force: since February 2025, the AI Act prohibits emotion recognition systems in the workplace, including hiring, except for narrow medical and safety purposes. Any vendor still selling facial-expression or emotion-based candidate scoring in the EU is selling a banned practice. Ask vendors directly what signals their scoring uses, and get the answer in writing.
How Does GDPR Apply to AI-Scored Interviews?
GDPR ran ahead of the AI Act by six years, and its rules still do most of the day-to-day work:
- Article 22: Candidates have the right not to be subject to decisions based solely on automated processing that significantly affect them. Rejecting a candidate is such a decision. The fix is meaningful human review before any decision, not a rubber stamp. See the full text at Article 22 GDPR.
- DPIA: Systematic evaluation of candidates using new technology typically triggers a Data Protection Impact Assessment under Article 35. Do it before rollout, not after a complaint.
- Transparency and lawful basis: Tell candidates what data you process, why, and for how long. Legitimate interest generally covers recruitment processing, but the balancing test must be documented.
- Data minimization and retention: Record what the assessment needs, nothing more, and delete recordings on a defined schedule. Video is personal data, and under the right conditions it can reveal special-category data, so retention discipline matters.
- Candidate rights: Be ready to honor access, erasure, and objection requests, including for interview recordings and scores.
A Practical Fairness Checklist for Hiring Teams
- Define competencies per role, then write structured questions that target them.
- Use identical questions, time limits, and conditions for every candidate at the same stage.
- Confirm your vendor's scoring uses answer content only, with no facial or emotion analysis.
- Inform candidates that AI-assisted scoring is used and what happens with their data.
- Keep a human decision before every rejection or advance, and log overrides.
- Run adverse impact checks on pass-through rates at least quarterly for high-volume roles.
- Complete a DPIA and set retention periods before the first interview goes out.
Fairer Screening in Practice
This is the model StormInterview is built around. Candidates complete async video interviews with identical structured questions on their own schedule. AI scoring evaluates each response against the role's rubric and presents evidence, not verdicts. Recruiters then make every call in the swipe-review workflow, screening 200 candidates in about 30 minutes with a full audit trail behind each decision. The result is a process that is faster than CV screening and, more importantly, more consistent and more defensible.
Start a free trial of StormInterview and replace 7-second CV scans with structured, human-controlled, AI-assisted screening your compliance team can stand behind.
Frequently Asked Questions
Is it legal to use AI in hiring in the EU?
Yes. The EU AI Act classifies hiring AI as high-risk, which means it is permitted with obligations: human oversight, transparency to candidates, monitoring, and logging. The main prohibition is emotion recognition in the workplace, which has been banned since February 2025.
Can AI make the final hiring decision?
No. GDPR Article 22 gives candidates the right not to be subject to solely automated decisions with significant effects. AI can score and rank; a human must meaningfully review and decide. Swipe-review workflows keep that human decision explicit and logged.
Does AI reduce hiring bias or make it worse?
It depends entirely on design. AI trained on biased outcomes replicates bias. AI that scores structured, job-relevant interview answers against a fixed rubric, validated with adverse impact testing, is measurably more consistent than manual CV screening, where identical resumes can see callback gaps of around 50% based on name alone.
What is adverse impact testing?
It compares selection rates across demographic groups. Under the widely used four-fifths rule, if one group's pass-through rate is below 80% of the highest group's rate, you investigate and remediate. Regular testing is also what regulators and bias-audit laws such as NYC Local Law 144 expect.
Do candidates need to be told AI is scoring their interview?
Yes. Both GDPR transparency requirements and the EU AI Act require informing candidates that AI is used in their evaluation. Clear disclosure also improves candidate trust and completion rates.