How Algorithms Can Reinforce and Amplify Society's Prejudices
By Marion Bordier
November 9, 2026
African-American English being disproportionately associated with negative stereotypes, facial recognition systems making more errors when identifying people of colour, and AI assistants offering safety advice simply because you mention being with an Algerian or Lebanese person: these examples raise a troubling question. Can artificial intelligence reproduce racial prejudice while presenting itself as neutral?
Speaking to Belgian broadcaster BX1 last June, Patrick Charlier, director of Unia, Belgium’s independent public institution for combating discrimination and promoting equality, pointed out: “Artificial intelligence can be racist.”
But for Hugues Bersini, professor of computer science at the Université libre de Bruxelles (ULB), an important distinction must be made, he told The Brussels Times.
“We cannot say that AI is racist because being racist involves taking a position or having negative feelings towards foreigners. AI has neither feelings nor intentions," he said. "It is still a reflection of our humanity, whether through the data (which is drawn from human beings) or through the algorithms (which are written by human beings)."
Yet AI can reproduce and amplify the prejudices present in the information it processes, turning existing stereotypes into seemingly authoritative answers.
AI models are trained on vast amounts of information, much of it taken from the internet. This material reflects society’s inequalities and stereotypes, as well as the people whose voices dominate online.
“The AI develops a certain conception of the world,” Bersini said. Because online content has historically been disproportionately produced by white men, AI-generated responses may reflect their perspectives more strongly than those of other groups.
Bias can arise from the way AI systems are developed. “Calling AI racist amounts to absolving software developers, who are the only ones truly responsible,” Bersini insisted.
The design of these systems must also be scrutinised. Élise Degrave, professor of digital law at the University of Namur, identified two main sources of bias: biased training data and the choices developers make when designing algorithms.
“Some of these criteria can lead to discrimination, even if they appear neutral,” she told The Brussels Times.
For example, an administration that instructs an algorithm to prioritise fraud detection in a poor neighbourhood may expose its residents to more inspections than people living in wealthier areas. Postal codes and names can also act as proxies for social background or ethnic origin, even when race is not explicitly considered.
When discrimination looks objective
The consequences become particularly serious when algorithms influence access to employment, housing or social benefits. “An algorithm can give these forms of discrimination the appearance of scientific objectivity,” Degrave warned.
Decisions that would be questioned if made by a human may be more readily accepted when presented as the result of an automated calculation.
Algorithms can also subject certain groups to repeated checks. Degrave calls this “algorithmic harassment": people are scrutinised without necessarily understanding why, even when they have done nothing wrong.
Meanwhile, the complexity and opacity of automated systems make discriminatory decisions difficult to identify and challenge.
For Nadine Brauns, who works on policy at Unia, compliance with the rules on transparency and explainability of systems is essential to protecting people against algorithmic discrimination.
Proving discrimination is already difficult, but automated systems make it even harder: it becomes more challenging to identify the criteria that influenced a decision or to understand how it was made.
Unia has so far received relatively few complaints explicitly identified as algorithmic discrimination, although reports linked to AI are beginning to emerge in employment, Brauns said. The organisation expects more cases as automated systems become widespread, potentially affecting hundreds or thousands of people at once.
In 2025, Unia opened 757 files concerning racial discrimination, a marked increase after several years of decline. Many concerned employment, alongside hate speech, hate crimes and online media. These figures relate to racial discrimination generally, not specifically to AI.
Can Europe regulate the problem?
The EU’s Artificial Intelligence Act introduces safeguards for certain high-risk systems, including those used in employment, education and public services. These systems face stricter requirements intended to protect fundamental rights.
But regulation is not a complete solution. Degrave pointed out that conventional algorithms can also discriminate without falling within the Act’s definition of AI. She advocates broader safeguards, including preliminary bias testing, independent audits and the involvement of lawyers, ethicists and other non-technical experts in system design.
Responsibility must also remain human. AI has no legal personality: accountability lies with the developers, organisations deploying the technology and decision-makers relying on its results. People must be able to understand how decisions affecting them were made and challenge them when necessary.
Another concern, Bersini warned, is that AI systems increasingly risk learning from content generated by AI itself. As machine-produced material spreads online, future models may absorb and reinforce the biases of previous systems, creating a feedback loop.
Generative AI can also be deliberately prompted to produce racist content or manipulated images. Its ability to create convincing material raises the risk of political disinformation, including anti-immigration propaganda.
For Degrave, tackling these problems requires more than technical adjustments.
“The design of algorithmic systems should not be left solely to computer scientists or engineers,” she argued. Citizens, public officials, legal experts and civil society must also be involved in assessing their consequences.
Copyright © 2026 The Brussels Times. All Rights Reserved.
