AI w rekrutacji — kiedy algorytm odrzuca kandydatów niesprawiedliwie
Automatyczne odrzucanie
Wielu z nas spotkało się z sytuacją, w której zgłoszenie znika bez śladu po wysłaniu na ofertę pracy. Coraz częściej za tym stoi algorytm — system analizuje tysiące aplikacji w sekundy, wyodrębnia kluczowe słowa, ocenia kandydatów i odrzuca dolną część przed tym, nim ktokolwiek je zobaczysz. Amazon, Goldman Sachs, Unilever — znane firmy, które w pewnym momencie korzystały z takich narzędzi. Problem nie jest w prędkości. Problem jest w tym, czego algorytm się nauczył.
Uczenie się uprzedzeń
Badanie opublikowane w 2024 roku przez zespół z University of Washington testowało, how three large language models (Mistral AI, Salesforce, Contextual AI) rank real CV with identical qualifications but different names. Result: names associated with white candidates were preferred in 85% of cases. Female-associated names — only 11%. Black male-associated names were never preferred over white male-associated names. The algorithm did not invent these prejudices — it learned them from historical data, in which more women applying for engineering roles were rejected, and more Black candidates applying for the same roles were rejected. The algorithm treated historical rejection rates as a signal of candidate quality. It was not programmed to discriminate. It was trained to optimize for “success” — and it learned that “success” looked a lot like the people already hired.
Skala problemu
A Stanford HAI study analyzing 4 million applications across 150+ employers using the same AI platform found that 26% of Black applicants and 15% of Asian applicants applied to positions where the system showed disparate impact — meaning a specific group is systematically treated worse, even if intentions are pure. In 2023, the US Equal Employment Opportunity Commission (EEOC) issued guidance warning that using algorithms in recruitment can violate Title VII of the Civil Rights Act. That same year, EEOC won a case against iTutorGroup — the first federal case where AI automatically rejected female applicants over 55 and male applicants over 60. The case settled for $365,000. In 2025, a California court certified collective action in Mobley v. Workday — hundreds of millions of applications, potentially making it the largest AI-hiring class action in US history.
Co robić
Firmy must conduct bias audits in their recruitment algorithms — regularly, not just at deployment. Tools like AI Fairness 360 (from IBM) or What-If Tool (from Google) help test whether the system treats equal groups equally. But even these tools are not enough — you must also audit the training data. And here lies the key: an algorithm cannot be better than the data used to train it. The European Union is responding — recruitment algorithms are classified as high-risk in Annex III of the EU AI Act, which mandates audits and transparency. Poland is still waiting for implementation.
Oznaczeniecontents AI
Ten powstał z participation of artificial intelligence: research and text were prepared and edited by an AI agent. Data, dates, and quotes cited in the text come from verified sources.
