At a time when 40% of jobs are exposed to Artificial Intelligence, and its use is becoming increasingly widespread, it’s worth asking the question of AI’s supposed neutrality: is AI really devoid of bias? A recent study by the Berkeley Haas Center for Equity, Gender and Leadership analyzed 133 Artificial Intelligence systems in different business sectors. The result: 44% of this panel displayed sexist stereotypes. So how do we explain this? Here’s how.

Where do these biases come from?

As Zinnya del Villar, Responsible AI expert at UN Women, explains, AI is first and foremost a question of data. To develop machine learning models, for example, or even generative AI tools, we give the algorithm data on which it can train. However, the data on which AIs are trained is mainly human-based data. Or rather, on all the data amassed there are real blind spots on certain segments of society: women, racialized people(according to the Berkeley Haas Center study, 25% of the AIs studied contained both sexist and racist stereotypes).

What’s more, the lack of diversity in the tech professions, and particularly in AI, also has something to do with it. Indeed, according to one study, 88% of AI algorithms are developed by men, and women account for just 30% of those working in the AI field. The people employed today to develop and maintain AI systems are predominantly white, male academics. Yet human influence can hardly be eliminated from data, since in most cases it is humans (and therefore men) who decide how data should be collected and then how it will be categorized and moderated.

Secondly, when AI learns to make sense of human language, it learns not only to account for the meaning of words, but also to understand which words are often associated together. This method of learning is fertile ground for the perpetuation of gender bias and essentialization processes. Clearly, if you ask an AI to generate a doctor and a nurse, it will tend to generate a male doctor and a female nurse.

Finally, Hélène Molinier, UN Women’s advisor on gender equality and digital cooperation, points out that there is currently no system in place to supervise or regulate the AI market. There is nothing to prevent the creation and use of AIs that reproduce sexist and/or racist stereotypes, or that fail to comply with confidentiality and security standards. As for taking into account the new forms of social vulnerability generated by AI, this is still in its infancy.

Biased AI, consequences for gender equality:

Biased algorithms are not neutral. According to Genevieve Smith and Ishita Rustagi, the authors of the Berkeley Haas Center study, ” Artificial Intelligence automates judgments that were previously made by individuals or groups of individuals “. AIs are mirrors of society: they integrate and automate the biases that prevail in it, and thus tend to amplify them.

This trend can be seen in a number of fields, including healthcare. As AIs are generally trained on male symptoms, they tend to propose the wrong diagnoses and treatments for women. Furthermore, Zinnya de Villar also points out that voice assistants, which often use female voices, tend to reinforce the association between women and care and service work. AI can also limit women’s professional opportunities in areas of decision-making, loan approval, recruitment or even judicial decisions.[Utilisate1] In 2018, for example, Amazon disabled an AI recruitment tool that favored male resumes. Indeed, if the data the AI is working with is biased, it will tend to associate men and women with certain roles and not others (we can take the example of doctors and nurses again).

But that’s not all. AI could also act as a brake on women’s employability. It could lead to a major restructuring of professions with a high potential for automation, particularly administrative and clerical functions. While AI tools can be springboards in terms of productivity and sometimes careers for those who use them, they also tend to widen the gap and marginalize those who don’t. In fact, women use AI tools around 20% less than men, both because of a lack of time and training, and because of self-censorship linked to a feeling of illegitimacy when it comes to anything to do with new technologies.

How can we make AI a vector for professional equality?

AI is malleable. It therefore has the capacity to be a real crutch in reducing inequalities, offering women access to new resources and opportunities through access to information and training. But for this to happen, gender equality needs to be taken into account right from the design and construction of AI systems.

This means actively selecting data that takes into account the diversity of social backgrounds, cultures, genders and roles. It also means getting to the root of the problem and diversifying AI systems development teams to make them more inclusive. Indeed, while 73% of executives believe it is important to have more women in this field, only 33% of them have a woman in a position leading strategic decisions on the subject[Utilisate2]. It is also essential to bring in gender expertise at the design stage to internalize vigilance for gender bias into the design process.

Finally, it is crucial to democratize the training and use of AI tools among women, so that they can take full advantage of the productivity and learning opportunities they offer, rather than falling victim to increasing marginalization.