> This is the markdown version of https://www.classet.ai/blog/ai-only-recruiting-more-biased-than-combined > Learn more at https://www.classet.ai ![AI-only recruiting bias compared to AI and human recruiting combined](/_next/image?url=%2Fimages%2Fblog%2Fai-only-recruiting-more-biased-than-combined.png&w=3840&q=75) # New Study: AI-Only Recruiting Is More Biased Than AI and Human Recruiting Combined June 2026 study shows AI-only recruiting creates more bias than AI-human hybrid recruiting. Learn why human oversight with clean AI beats automation alone. [![Paul Jones](/_next/image?url=https%3A%2F%2Fassets.basehub.com%2Fe0b5701f%2F6599306507912123f90f150a8bfaaf6c%2Fscreenshot-2026-01-28-at-10.53.16-am.png%3Fwidth%3D100%26height%3D100%26quality%3D100&w=96&q=75) Paul JonesHead of Growth at Classet ](/blog/authors/paul-jones) June 15, 2026 AI Recruiting, Guides & Insights Most TA teams running [AI recruiting bias](/blog/ai-recruiting-bias-audits-compliant-platform) audits assume human oversight fixes the problem. A University of Washington study in 2025 proved otherwise. When recruiters reviewed candidates using biased AI, they followed the AI's inequitable choices up to 90% of the time. Without the AI or with an unbiased model, those same recruiters selected candidates fairly. The lesson is clear: biased AI screens candidates wrong and reshapes how humans screen them too. **TLDR:** - Recruiters mirror biased AI recommendations up to 90% of the time, per 2025 University of Washington study - Amazon scrapped its AI recruiting tool in 2018 after it penalized resumes containing "women's" - Ninth Circuit ruling lets AI vendors face liability under anti-discrimination law, not just employers - NYC Local Law 144 requires annual third-party bias audits before deploying automated screening tools - Classet's Joy conducts phone screens but never makes hiring decisions; recruiters always decide ## What AI Recruiting Bias Is and Why It Matters in 2026 AI recruiting bias refers to patterns where automated hiring systems produce outcomes that disadvantage candidates based on race, gender, age, or other protected characteristics, often without any explicit intent to discriminate. These patterns arise when AI systems learn from historical hiring data that already reflects human bias, or when [the training data underrepresents certain groups](/blog/ai-recruiting-vendor-questions). The stakes are real. A large share of employers now use some form of automated screening, yet many lack the internal audit processes to catch discriminatory outputs before they affect real candidates. Bias in hiring harms applicants and exposes organizations to growing legal liability as regulators catch up with the tech. What makes this hard to fix is that AI bias often goes undetected longer than human bias. A recruiter's bad judgment is visible in one decision. A biased algorithm quietly shapes thousands. ## Real Examples of AI Bias in Hiring Three cases stand out as the clearest real-world evidence that AI hiring tools can encode and amplify bias at scale. ### Amazon's Recruiting Tool (2018) Amazon built an AI recruiting tool trained on a decade of resumes. Because most hires during that period were men, the model learned to penalize resumes containing the word "women's" and downgraded graduates of all-women's colleges. Amazon scrapped the tool in 2018 after finding it could not be reliably fixed. ### Workday Lawsuit A class-action suit filed against Workday alleged its AI screening tools discriminated against applicants based on race, age, and disability. The case raised serious questions about vendor liability when automated candidate screening produces disparate outcomes across protected groups. ### HireVue and Facial Analysis HireVue faced scrutiny over video interview software that scored candidates partly on facial expressions and tone. Critics argued this approach introduced bias against candidates with disabilities and non-native speakers, leading the company to remove facial analysis from its product. These cases share a common thread: when AI screens candidates without meaningful human review, bias in the training data becomes bias in the outcome. Each one is now cited in regulatory discussions around AI hiring software and automated candidate screening bias. ## The Workday Lawsuit: A Landmark Case for AI Hiring Liability In 2023, Derek Mobley filed a class action lawsuit against Workday, alleging its AI-powered screening tools discriminated against him based on race, age, and disability. He claimed to have been rejected by over 100 employers using Workday's applicant tracking system. The case marked [the first vendor liability ruling](/blog/ai-hiring-under-scrutiny-the-eightfold-lawsuit-and-what-it-means-for-your-vendor-selection), with a vendor held directly liable for discriminatory hiring outcomes alongside the employer. A federal judge initially dismissed the case in 2024, but the Ninth Circuit Court of Appeals revived it in early 2025, ruling Workday could face liability as an employment agency or agent. That ruling set a precedent: AI hiring software vendors can be held accountable under federal anti-discrimination law, and employers who buy their tools are no longer the only ones on the hook. ### Why This Case Matters for TA Leaders The Workday lawsuit signals a shift in how courts may treat automated candidate screening. If liability extends to vendors, it changes the risk calculus for any company using AI hiring software without human review built into the process. Questions about how screening decisions get made, and who is accountable for them, are no longer hypothetical. ## How AI Learns to Discriminate AI hiring tools learn from historical data, and that's where the problem starts. When past hiring decisions reflected human bias, those patterns get encoded into the model. The system mirrors whoever made the decisions before it. Three core mechanisms drive this: - Training data bias occurs when the data fed to a model reflects historical inequities. If past hires skewed toward one demographic, the model learns to favor that group. - Proxy variable bias happens when seemingly neutral inputs like zip code, school name, or employment gaps link to protected characteristics, allowing discrimination to persist without explicit intent. - Feedback loop bias compounds over time. When a biased model makes decisions that generate new hiring data, that data gets fed back into future training, reinforcing the original bias with each cycle. ### Why Auditing Alone Doesn't Fix It Many organizations treat a one-time bias audit as a solution. But if the underlying training data hasn't changed, audits only confirm what's already broken. [Research shows](https://www.science.org/doi/10.1126/science.aal4230) that even well-intentioned AI systems can perpetuate discrimination when the feedback loop goes unchecked. The fix requires both cleaner inputs and human judgment applied at the point of decision, before outcomes ever reach candidates. ## University of Washington Study: Humans Mirror AI Bias Up to 90% of the Time A 2025 University of Washington study tested the "human in the loop" assumption directly, and the results complicate the narrative. When recruiters reviewed candidates using AI tools with bias baked into the models, they mirrored the AI's inequitable choices up to 90% of the time. The AI's judgment effectively became their judgment. Yet when those same recruiters made decisions without any AI assistance, or with an unbiased AI, they selected white and non-white candidates at equal rates. > The problem goes beyond biased AI. It's that biased AI reshapes human judgment too. Human review only works as a safeguard when the AI feeding it is clean. A biased model steers the humans reviewing its outputs toward biased decisions. Oversight of a flawed system is not the same as oversight that actually catches flaws. ## The Legal Framework: NYC Local Law 144, California, and State-by-State Requirements A patchwork of state and local laws is making AI hiring compliance non-negotiable. ![Map of US states with AI hiring regulations, shaded by regulatory intensity](https://d4bkhhmrfehmf.cloudfront.net/media/46c5fb95-2bc3-485e-94d2-7604fb1bce97/TwUIHvW9yQdDe5FrD2jLd.png) Jurisdiction Law Core Requirement New York City Local Law 144 Annual third-party bias audit + public disclosure before deploying AEDTs California Civil Rights Council (Oct 2025) Disparate impact liability extended to algorithmic screening tools Illinois HB 3773 Candidate notice required when AI is used in evaluation Colorado AI Act Risk assessments required for high-risk AI systems Texas TRAIGA Accountability requirements for AI developers and deployers NYC Local Law 144 sets the highest bar: conduct an annual third-party bias audit, publish the results publicly, and complete both steps before deploying any automated employment decision tool. California's October 2025 rules are the broadest in scope, applying disparate impact liability to any screening tool regardless of intent. For employers operating across multiple states, building to NYC's standard is the practical path to coverage. It won't satisfy every local nuance, but it covers the most demanding requirements on the map. ## Why Fully Automated Hiring Creates More Risk Than AI-Assisted Processes Fully automated screening carries the highest regulatory risk. When no human touches a decision before it's acted on, the system becomes the decision-maker, and that's exactly where courts draw the line. Human-in-the-loop has become the compliance floor because it preserves accountability. As the UW study showed, it doesn't automatically correct bias. But it keeps a person responsible for outcomes, and under NYC Local Law 144 and the frameworks building on it, that distinction is what regulators look for. Full automation compounds risk in two directions: biased outputs reach candidates with no catch point, and liability becomes harder to defend. Adding human review doesn't solve everything. Removing it, though, removes the last checkpoint before a bad decision becomes a pattern. ## Classet's Approach: Configurable AI Screening with Human Final Decisions Every issue covered in this post shaped how we built Classet. Joy conducts structured phone screens but never makes a hiring decision. Joy never makes suggestions on who to hire or not, but rather speeds the important information to recruiters, who receive a candidate summary covering relevant experience, availability, and motivations. The final call is always theirs. Candidates know who they're talking to. Every outreach interaction discloses upfront that Joy is an AI, and opting out is always available without affecting hiring eligibility. For ongoing compliance, we partner with Warden AI for continuous third-party bias monitoring, with [monthly results published publicly](https://trust.warden-ai.com/classet/ai-phone-interviewer). That's the same standard NYC Local Law 144 requires, applied continuously instead of only once a year. In [10,167 post-interview ratings](/land/candidate-experience), 88% rated their experience with Classet's Joy positively. Consistent AI screening paired with human final decisions produces better outcomes for both sides of the process. ## Fixing AI Discrimination Takes More Than an Annual Audit Fixing AI discrimination in hiring takes more than good intentions or one-time audits. When biased models quietly shape thousands of decisions, the damage compounds before anyone notices. You need clean training data, continuous monitoring, and real human authority over final outcomes. The Workday case proved vendors can be held liable, not just employers. If your AI makes hiring decisions without meaningful review, you're carrying regulatory and legal risk that courts are actively enforcing right now. ## FAQ Does AI reduce hiring bias compared to human-only screening? Not when used alone. A 2025 University of Washington study found that recruiters using biased AI tools mirrored the AI's discriminatory choices up to 90% of the time, but made equitable decisions when reviewing candidates without AI or with unbiased AI. Keep the data feeding those decisions clean and keep humans on the final hiring call. What happened with the Amazon AI recruiting tool bias in 2018? Amazon built an AI recruiting tool trained on a decade of resumes, but because most past hires were men, the model learned to penalize resumes containing "women's" and downgraded graduates of all-women's colleges. Amazon scrapped the tool after discovering the bias couldn't be reliably fixed, making it one of the clearest examples of how AI bias in hiring emerges from biased training data. Workday AI lawsuit vs other AI discrimination cases: what's different? The Workday case is the first where courts ruled an AI vendor, not just the employer, could face liability as an employment agency under federal anti-discrimination law. The Ninth Circuit's 2025 ruling means companies selling AI hiring software can be held accountable for discriminatory outcomes, fundamentally changing vendor risk and accountability for automated candidate screening bias. Can you build compliant AI hiring without human final decisions? No. Fully automated screening carries the highest regulatory risk because when no human touches a decision before action, the system becomes the decision-maker, exactly where courts and regulators like NYC Local Law 144 draw the line. Human-in-the-loop doesn't automatically correct bias, but it preserves accountability and creates a compliance checkpoint before decisions reach candidates. How does Classet handle AI bias in phone screening? Joy conducts structured phone screens but never makes hiring decisions. Recruiters receive candidate summaries and always make the final call. Every interaction discloses upfront that Joy is AI with opt-out available, and Classet partners with Warden AI for continuous third-party bias monitoring with monthly results published publicly at the same standard NYC Local Law 144 requires. What are the biggest AI bias types to watch for in recruiting tools? The three main types are training data bias (when historical hiring patterns reflect past discrimination), proxy variable bias (when neutral inputs like zip code correlate with protected characteristics), and feedback loop bias (when biased decisions create new training data that reinforces the original bias). Each type can persist even after a one-time audit if the underlying data and decision process don't change. Best way to handle AI recruiting compliance without changing our ATS? Look for AI screening tools that integrate directly with your existing ATS and keep humans in the final decision loop. Classet's ATS Sync connects to 100+ systems without workflow changes, and Joy presents candidate summaries while recruiters always make hiring decisions. Pair that with continuous third-party bias monitoring to stay ahead of regulations like NYC Local Law 144. Can AI phone screening verify certifications like EPA or CDL during the call? Yes. AI phone screening can ask candidates about EPA certification, CDL status, journeyman licenses, or state-specific credentials during the structured interview. Joy captures those responses in the candidate summary, but recruiters still verify documentation before making offers. When does it make sense to use AI screening vs hiring another recruiter? AI screening makes sense when you're hiring high-volume roles and need consistent top-of-funnel coverage without adding $90K–$130K in recruiter headcount. If your team spends most of their time on repetitive phone screens instead of closing qualified candidates, a configurable AI recruiter like Joy handles that work at a fraction of the cost while your recruiters focus on relationship-building. How do you join the Workday AI bias lawsuit? The Workday lawsuit is a class action case, so joining typically requires contacting the plaintiff's legal team or waiting for formal class certification notices. If you believe you were discriminated against by Workday's AI screening tools, consult an employment attorney who can connect you with the case representatives and advise on your specific situation. What's the difference between automated candidate screening bias and human hiring bias? Automated candidate screening bias happens when AI tools learn from biased historical data or use proxy variables that correlate with protected characteristics, often affecting thousands of candidates at scale. Human hiring bias is inconsistent and varies by individual recruiter. The University of Washington study showed biased AI can actually reshape human judgment, with recruiters mirroring the AI's discriminatory choices up to 90% of the time. Does human in the loop recruiting actually reduce AI discrimination in hiring? Only if the AI feeding the human is unbiased. The 2025 University of Washington study found that recruiters using biased AI mirrored its discriminatory choices 90% of the time, but made fair decisions when reviewing candidates without AI or with clean models. Human oversight catches flaws only when the underlying AI isn't steering them toward biased outcomes. Real life examples of AI bias beyond Amazon's 2018 recruiting tool? The Workday lawsuit alleges AI screening discriminated based on race, age, and disability across 100+ job applications. HireVue faced scrutiny for scoring candidates on facial expressions and tone, which critics argued biased against people with disabilities and non-native speakers. Each case shows how AI hiring bias examples emerge when training data reflects historical inequities or uses proxy variables tied to protected characteristics. AI hiring lawsuit trends: what changed after the Workday case? The Ninth Circuit's 2025 ruling in the Workday case established that AI vendors, not just employers, can face liability as employment agencies under federal anti-discrimination law. This shifts risk to software companies selling automated screening tools and raises the compliance bar for any organization using AI hiring software without meaningful human review built into the process. Why does AI recruiting bias get worse without continuous monitoring? Feedback loop bias compounds over time. When a biased AI makes hiring decisions that generate new data, that data gets fed back into future training cycles, reinforcing the original discrimination with each iteration. One-time audits only confirm what's already broken. Continuous third-party monitoring catches bias drift as new data enters the system and hiring patterns evolve. ## Ready to See Compliant AI Screening in Action? If you're comparing AI hiring tools and need to know your vendor can stand up to bias audits, [schedule a demo](/demo) to see how Joy handles structured phone screens while keeping humans in control of hiring decisions. Or review our [public bias monitoring results](https://trust.warden-ai.com/classet/ai-phone-interviewer) to see exactly how we measure fairness month over month. [![Paul Jones](/_next/image?url=https%3A%2F%2Fassets.basehub.com%2Fe0b5701f%2F6599306507912123f90f150a8bfaaf6c%2Fscreenshot-2026-01-28-at-10.53.16-am.png%3Fwidth%3D100%26height%3D100%26quality%3D100&w=128&q=75) Paul Jones Head of Growth at Classet Paul comes from an operator background running an Alpine-owned company, and brings firsthand experience with the hiring challenges Classet was built to solve. He's driven by a belief that the right technology can make meaningful work more accessible. ](/blog/authors/paul-jones) Follow Classet in Google Add classet.ai as a preferred source and our hiring research is more likely to show up in your Top Stories and AI Overviews. [Add as preferred source →](https://www.google.com/preferences/source?q=classet.ai) ## Explore More ### Use Cases - [RPO / BPO Recruiting](/use-cases/call-centers-bpo) - [Healthcare Recruiting](/use-cases/healthcare) - [Hospitality Recruiting](/use-cases/hospitality) ### Integrations - [Greenhouse](/integrations/greenhouse) - [Bullhorn](/integrations/bullhorn) - [Lever](/integrations/lever)