The use of artificial intelligence (AI) in hiring processes does not automatically eliminate bias, and concerns are growing about the data used to train these systems. As the European Union enforces new rules on AI, companies using AI in recruitment are facing closer scrutiny. Recent research suggests that algorithmic bias often doesn't come from the AI itself, but from human decisions embedded in the data used to train these systems. This means that even when AI systems are not explicitly programmed to discriminate, they can still produce unequal outcomes. In late 2025, a French government agency responsible for preventing discrimination raised concerns about how a major social media platform distributed job ads. The platform showed some ads more frequently to men and others to women. This case highlights a growing concern for both companies and policymakers: how can AI systems produce unequal results without being explicitly programmed to do so? This issue has become even more pressing with the enforcement of the European Union's AI regulation, which categorizes AI used in hiring and employment as high-risk. This classification makes it clear that simply removing sensitive data such as gender or ethnicity from AI systems is not enough to ensure fairness in hiring. Companies using AI to select or recommend candidates must look beyond basic data filtering to address deeper issues of equity. A common belief is that bias in AI only occurs when algorithms directly use protected characteristics like race or gender. However, the issue is often more complex. AI systems learn from historical data, which may already reflect existing inequalities. Many online job platforms and recruitment tools now use ranking systems that prioritize factors like past performance, number of completed tasks, or experience. While this approach seems logical for reducing hiring costs and identifying strong candidates, it can also reinforce existing inequalities. For instance, someone who starts with a slight career advantage may accumulate more experience, improving their ranking in automated systems and increasing their visibility for new opportunities. This creates a cycle where initial advantages lead to more opportunities, which in turn reinforce those advantages. As a result, a ranking that appears "objective" or "based on performance" can actually reproduce and amplify existing inequalities. Research on large online work platforms shows that these mechanisms can affect multiple groups at once. On average, women tend to complete fewer tasks than men with similar characteristics like age, education, and pay rate. This leads to lower visibility in rankings. The gap can widen further when multiple factors combine, such as being a woman from an ethnic minority, who may face compounded disadvantages. Similarly, young professionals, who often have less experience to highlight, may be systematically ranked lower despite having comparable skills. This "visibility deficit" doesn’t necessarily reflect their actual abilities but highlights how structural disadvantages can be reinforced through automated systems. For companies using AI in hiring, talent management, or human resources, several key lessons emerge. First, removing sensitive data like gender or ethnicity is not enough to prevent bias, as systems can still learn from biased historical data. Second, the criteria used to determine the "best candidates" must be carefully evaluated, as they can unintentionally favor those who have already had more visibility or access to opportunities. Finally, feedback loops—where AI systems influence candidate visibility and that visibility shapes future data—must be considered a major governance challenge. The new European AI regulations signal a shift in perspective. The focus is no longer just on whether an algorithm uses sensitive data, but also on whether it can produce discriminatory effects through indirect mechanisms. These challenges extend beyond hiring, affecting areas like credit approval, insurance pricing, personalized recommendations, and content moderation. Understanding the origins of the data used to train AI systems is now as important as the technical design of the models themselves.