Is generative Artificial Intelligence sexist?

Generative Artificial Intelligence (AI) is becoming an increasingly common part of our daily lives. It’s at the heart of the chatbots that answer our calls and the bots that assist us on websites, and it’s the star of image and text generators, including the now-famous ChatGPT.

Its rapid rise in popularity—whether in our daily lives or at work—is bound to raise a host of questions. The most common one is: “Will it take our jobs?” Next comes: “Is it reliable?”, which can be interpreted as “Does it tell the truth?”. This concerns not only the trust we can place in it to perform the tasks we intend to assign to it, but also its impact on mindsets and social perceptions—and, consequently, the consequences that generative Artificial Intelligence may have on reality, social relationships, inequalities, and so on.

This issue has been explored internationally by UNESCO, which, in its report “ Challenging Systematic Prejudices: an investigation into bias against women and girls in large language models ,” raises the alarm. The three generative AI models studied by UNESCO researchers produce up to 20% sexist, even frankly misogynistic, content. How does this manifest itself? Through the association of terms such as “family,” “home,” “child,” “emotion,” and “assistant” (and other more directly hostile or degrading ones) with female representations, and the association of terms such as “salary,” “business,” “success,” and “management” with male representations.

But why is that? Because generative Artificial Intelligence “learns” only from existing content. However, if that content contains biases, the Artificial Intelligence simply internalizes them… But where this becomes critical is that the “training” of Artificial Intelligence—whose very principle is selection—will amplify these errors. In fact, it is nothing more than a selection bias that we humans also possess. Our brains spend their time “sorting” the information that comes to them and organizing it according to what seems most useful for our survival, most conducive to our comfort, and most efficient in terms of effort.

The difference is that we have a conscience, and that conscience is influenced by our consideration of the real-world consequences of our decisions. So, for example, when we associate the idea of management with the image of a man, we are able to realize that, on the one hand, this does not reflect reality and, on the other hand, it could discourage women from pursuing their ambitions. Generative Artificial Intelligence, on the other hand, has no contact with reality, but only with its representations. We cannot, therefore, ask it to correct itself.

But then, how can we stop the cycle of sexism generated by Artificial Intelligence? The solution most consistent with how generative Artificial Intelligence works would be to train chatbots and other applications that incorporate this type of algorithm. The idea would be to encourage the Artificial Intelligence to correct itself by exposing it to a wider variety of non-stereotypical sources. But researchers at UNESCO are pessimistic about the effectiveness of this approach: Generative Artificial Intelligence is being deployed too quickly compared to our human capacity to provide it with non-sexist material, even as it has an abundance of biased sources at its disposal. They therefore recommend pursuing a path of regulation and judicial intervention. In other words: holding the companies that produce and operate generative Artificial Intelligence models responsible for verifying that the generated content is not discriminatory and for correcting it if necessary. In the event of a breach of these obligations, these companies could be prosecuted and penalized.

So far, no international governing body or national parliament has been able to pass legislation along these lines. So, in the meantime, let’s remain vigilant in our use of chatbots: they can sometimes be even more biased than we are!