Often perceived as neutral and objective, algorithms are nonetheless an increasingly prominent part of our daily online lives. Recruitment, advertising, content recommendations… Their decisions influence our opportunities and our perceptions of the world. But are these systems truly impartial? This is the context for the work of Grazia Cecere, a researcher at Institut Mines-Télécom Business School. On the occasion of International Women’s Day, she analyzes the mechanisms of algorithmic gender bias, their societal impacts, and possible solutions.
The Illusion of Algorithmic Neutrality
Algorithms enjoy a reputation for objectivity: based on mathematical and statistical models, they appear to make rational decisions, in contrast to human judgments. Yet this neutrality is illusory.
In 2023, a Bloomberg investigation revealed that the image-generating AI Stable Diffusion, developed by Stability AI, produced highly stereotypical representations of professions: architects who were predominantly white and male, cleaning staff who were mainly Black… Behind the technical performance, social biases persist.
Algorithms learn from data produced by our societies. When this data reflects existing inequalities (such as the underrepresentation of women in certain professions), artificial intelligence systems tend to incorporate them. This is what Grazia Cecere calls the mirror effect: the algorithm reproduces the stereotypes already present in the real world.
When Technology Creates New Forms of Discrimination
Algorithms do more than simply replicate existing inequalities; they can also generate new ones.
In online advertising, for example, certain job postings for high-level positions are shown less frequently to women. This is not a deliberate exclusion: advertising systems simply calculate that these audiences are more expensive to reach. Algorithms therefore optimize ad delivery to the targets deemed most “profitable.”
This economic logic produces indirect discrimination: no explicit human intent, but a result that reinforces inequalities.
Where do these biases come from?
Grazia Cecere’s research identifies several sources: human choices and data limitations.
Regarding human biases, the study explains that algorithms are designed by teams of engineers (who are, of course, human) whose design choices reflect cultural and cognitive frameworks. The way individuals are categorized can lead to exclusions or overrepresentations.
As for data biases, incomplete or non-representative datasets contribute to the reproduction of inequalities. For example, the predominance of men in engineering programs can lead a system to automatically associate that field with a particular gender.
A revealing experiment on visual stereotypes
To analyze these mechanisms, Grazia Cecere conducted an experiment on the dissemination of advertising images. The study draws inspiration from critiques of the portrayal of women in film, particularly the “Headless Women of Hollywood” project, which condemns the dehumanization of women on screen. The researchers tested photographs of men and women (fully clothed), some of which were cropped to exclude the face. The result: images of “headless” women were promoted more by the algorithms.
This experiment reveals a vicious cycle: users, already exposed to these visual stereotypes, interact more with these images. The algorithms, trained on this behavior, amplify their dissemination.
Auditing Algorithms: A Scientific Challenge
Identifying these biases requires the ability to observe how the systems work. However, many algorithms operate like black boxes, with no transparency regarding the code or the data used.
Researchers therefore turn to experimental methods:
- A/B testing: comparing groups exposed to different versions
- A/A testing: simultaneous comparison under strictly identical conditions, particularly suited to dynamic environments such as online advertising
These techniques make it possible to isolate the effects related to the algorithms’ actual operation.
Correcting biases: possible, but complex
Some biases can be mitigated by rebalancing the data, for example by manually adjusting the representation of underrepresented groups.
But when biases arise from the autonomous operation of systems, they become harder to anticipate. Content that generates more clicks (often stereotypical) risks being over-promoted. These dynamics are studied in research on “algorithmic collusion,” where decisions emerge without direct human intervention.
Legal frameworks are also evolving to address these challenges. The Digital Services Act now requires major digital platforms to take responsibility for the effects produced by their algorithms. Beyond legal obligations, reputational risks also encourage companies to take action. For researchers, the solution lies in close collaboration between academia and industry, the development of robust monitoring tools, and the implementation of technical safeguards.
Toward a More Equitable AI
According to Grazia Cecere, gender bias is a problem that can be technically resolved. Methods exist to correct imbalances when data is insufficiently representative.
Other forms of discrimination remain more difficult to measure. In France, for example, the ban on collecting ethnic data limits the identification of certain biases, unlike practices permitted in other countries such as the United States.
Transparency, continuous monitoring of systems, and raising awareness among the scientific and industrial communities therefore remain necessary to ensure the ethical development of artificial intelligence.