At the end of 2025, the French authority responsible for fighting discrimination raised concerns about the job posting distribution system of a major social media platform. Some advertisements were shown more frequently to men, while others were shown more frequently to women. This case highlights a growing question for companies and public officials: how can artificial intelligence (AI) systems produce unequal results without being explicitly programmed to discriminate? This issue has become more urgent with the European AI regulation, which classifies systems used in recruitment and employment as high-risk. The regulation makes it clear that simply removing sensitive variables like gender or ethnic origin from AI systems is not enough to ensure fairness in hiring and job recommendations. A common belief is that biases only occur when an algorithm directly uses protected characteristics such as race or gender. However, the problem is often more complex. AI systems learn from historical data, which may already reflect past inequalities. Many online work platforms and recruitment tools now rely on ranking systems that prioritize indicators such as past performance, the number of completed tasks, or accumulated experience. From a company's perspective, this logic seems rational because it reduces search costs, helps find relevant candidates, and rewards experience. However, these indicators can also reinforce existing inequalities by assuming that past success reflects individual merit rather than access to opportunities. This creates a feedback loop: someone who gets a small advantage early in their career may accumulate more experience, which improves their ranking in automated systems. This increased visibility can lead to more opportunities, further strengthening their advantage. For companies using AI-based recruitment tools, this means that a ranking labeled as "objective" or "based on performance" might 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 complete fewer tasks than men with similar characteristics, such as age, education, and hourly rate. This leads to lower visibility in rankings. The gap can be even larger when multiple factors combine—for example, women from ethnic minorities may face more disadvantages than expected based on each factor alone. Similarly, young professionals often have less experience to showcase and may be systematically disadvantaged in rankings, even if their skills are comparable. This results in a "visibility deficit" that doesn't necessarily reflect their actual abilities. For organizations using AI in recruitment, talent management, or human resources, several lessons are clear. Removing sensitive data like gender or ethnicity is not enough to prevent bias. A system can still reproduce inequalities if it learns from past decisions that were themselves biased. Even if sensitive data is removed from freelancer profiles, the initial human bias may still be embedded in the data. Optimization criteria also need careful review. Tools that identify the "best candidates" might favor those who have already had more visibility or access to opportunities. Finally, feedback loops—where a system influences candidate visibility, which in turn affects future data—should be treated as a serious governance challenge. The implementation of European AI regulation marks a shift in focus. The question is no longer only whether an algorithm uses sensitive data, but also whether it produces discriminatory effects through indirect mechanisms. These challenges extend beyond recruitment. Today, algorithms also influence credit granting, insurance pricing, personalized recommendations, and content moderation. In all these areas, understanding the origin of the data is as important as the technical design of the models.