2025年人才招聘领域人工智能偏见状况报告

State of Al Bias in Talent Acquisition 2025 A data-driven review of Al bias, compliance,and responsible Al practices in TA Foreword by Kyle Lagunas responsibly, I've seen excitement around Al quickly give way to concern,especially around bias and fairness. As Al tools embed deeper into hiring and workforce decisions, we face adefining moment: will we carry forward old inequalities, or use Al to buildsomething better? New and forthcoming legislation make this choice all the more urgent. Pendingcourt cases signal that Al risk is no longer theoretical. For HR and TA leaders,and technology vendors, bias is now a legal and reputational issue. we face.This report brings a number of interesting points together to crystallize this Founder & Principal,Kyle & Co critical conversation. It offers a grounded view of where Al is impacting TAWhile bias is real, it is also measurable, manageable-and, thankfully, mitigatable.MWardenAl Executive summary Bias in Al has become one of the most urgent and emotiveissues in hiring, with high-profile lawsuits like Mobley vs.Workday pushing the risks into the legal and media spotlight.This report cuts through the noise with a data-rich view of howbias is playing out in real-world Al systems. Vendors are investing in responsible Al practices and preparingto meet Al regulations, but many have a long way to go. There'sa promising level of transparency with the majority of vendorssharing publicly available documentation. Yet transparency toend-users is limited with only 15% made aware that Al is used. 75% of HR/TA leaders cite bias as one of the top concerns withadopting Al. At the same time, they are under pressure to solvebusiness problems using Al, creating a sharp tension. The question is no longer whether Al can be biased, but how wemanage its use in a way that improves outcomes. Done right,Al offers a real opportunity to reduce longstanding inequalities,but only if it's developed, chosen, and deployed with care. some Al systems do show bias, 85% of Al systems metaccepted fairness thresholds. 75% of HR leaders cite bias as a top concern whenevaluating Al tools.Al can be demonstrably biased The data suggests that Al systems can deliver up to 45% faireroutcomes for women and racial minorities than human-leddecisions.That said, bias does exist in some systems, 15% of tools failed 15% of Al systems fail standard fairness benchmarks andmany academic studies identify bias in lab settings.Yet Alis typically more fair than humans In some cases, Al is up to 45% fairer for womenandracial minorities.Vendors are investing in responsible Al practices basic fairness tests, and the gap between Al systems anddifferent vendors varies by up to 40%. Practices are evolving but transparency to end-users islacking.V Warden Al About this report This report analyzes high risk Al systems, such as those used toinfluence decisions in talent acquisition and intelligence, utilizingmultiple data sources: Bias audits from real systems 150+ audits of Al/automated tools using over 1 million testsamples, tested for fairness using standard methods. survey contributions 5o+ responses on Responsible Al practices, transparency,and regulatory readiness. HR tech vendors A blend of academic and industry studies on human bias, converted into comparable fairness metrics. Test samples ioo+ public documents from vendors sharing audits and transparency information. Special thanks to our contributors Censiahackajob 三 findem sapiaOJOBMA Table of contents How big of a concern is Al bias?06 Bias BenchmarksHow biased are Al systems in practice?14 Al vs. HumansHow fair is Al compared to humans?20 Is the industry ready for Al regulation?.27 How mature are vendors on Responsible Al?32Conclusion PerceptionHow big of a concern is Al bias? "Buyers are asking more questions about Albias and it's becoming a key factor inprocurement decisions." Chief Legal Officer,Greenhouse Software Perception TA is embracing Al more than ever before :: Three out of four TA teams say they're already using Al or automated systems. What this means about impact of Al solutions. PerceptionBut concerns over Al risks are catching-up o: Al risk is now a commercial issue, not just a technical one What this means There's a strong risk/reward trade-offwith Al solutions. Typically, the morevalue the Al delivers, the more it islikelytobeautomatingof apreviouslyperson-powered process, and themore risk there is from reducedhuman control. How often are Al risks akey buying factor? are waking up to the Al risks andadapting their procurement gaining traction in court (e.g. Mobley Bias is one of the top concerns slowing Al adoption 75% of HR leaders name bias as a top concern when evaluating Al,second only to data privacy What this means Data security and privacy are long-standing software risks that mostorganizations have become mature new phenomenon compared totraditionalsoftware. inevitably surround sensitive