Artificial intelligence learns from our data, but also from the choices of those who design it. Can it claim to understand all of humanity if it only reflects part of it?
We have long considered theartificial intelligence as a subject reserved for engineers. It has become a social issue.
Because AI no longer just executes our requests. It begins to guide our choices. It recommends medical treatment, suggests recruitment, prioritizes the information we consult, personalizes our education, influences our consumption and, tomorrow, will probably participate in some of the most important decisions of our lives.
A question then seems essential to me: who learns artificial intelligence?
The illusion of neutral technology
We talk a lot about the data that feeds the models; much less of those who design them.
However, artificial intelligence does not only learn from the cohorts of data provided to it. It also learns from the human choices that govern its design: what problems are worth solving? What criteria should be optimized? What risks are acceptable? What errors are tolerable? What objectives should be pursued?
None of these arbitrations are neutral.
We have often presented artificial intelligence as an objective technology. It is no more so than the human beings who develop it. Data bears the traces of our history. Algorithms reflect the priorities we assign to them. Models learn as much from our knowledge as from our limitations and our oversights.
When biases become visible
The first signals are already visible.
Facial recognition systems were found to be significantly less effective in identifying certain women, particularly those with dark skin. Recruitment tools have reproduced the biases of past hiring by disfavoring female candidates. Generative models still spontaneously associate scientific professions, managerial functions or technical professions with men, while they more easily link women to domestic tasks or care professions.
UNESCO has also shown that these stereotypes remain largely present in the most recent generative models.
These biases are not always the result of intention. They are often the consequence of an incomplete representation of the world.
The history of science warns us
The history of science reminds us how far this question is from being theoretical.
For decades, much medical research has relied primarily on male cohorts. The specific symptoms of certain cardiovascular pathologies in women were identified later. Diseases such as endometriosis have long remained insufficiently studied, delaying their diagnosis and treatment.
It is obviously not a question of opposing male medicine to female medicine.
It is a question of remembering that a science progresses when it broadens the way it looks at reality.
Why would it be any different with artificial intelligence?
Building a more universal intelligence
Today, women represent around 22% of AI professionals worldwide and only 12% of researchers specializing in this field. This underrepresentation is not just a question of professional equality. It questions the diversity of human experiences which contribute to the design of technologies intended to support billions of individuals.
An artificial intelligence intended to understand humanity cannot be built sustainably from a limited number of points of view.
The answer does not only consist of recruiting more women in tech, it must be promoted much earlier: by giving young girls a taste for mathematics, science and engineering from childhood, by making more visible the women scientists who, from Sophie Germain to Ada Lovelace, from Rosalind Franklin to Katherine Johnson, from Marie Curie to Emmy Noether, have profoundly transformed our understanding of the world.
By also promoting truly multidisciplinary teams where engineers, doctors, psychologists, philosophers, lawyers and creators interact and perhaps also by telling more stories.
Vocations are often born from stories that make us dream.
A question of civilization
For decades, cinema has enriched our understanding of the world by multiplying perspectives, sensitivities and human experiences. It is precisely this diversity that has allowed him to tell universal stories.
Artificial intelligence will face the same challenge.
We talk a lot about computing power, ever more efficient models or billions of additional parameters.
But the true wealth of an intelligence may not lie in the quantity of data it absorbs, but in its quality.
It lies in the diversity of perspectives which teach him to understand the world.
We will never build a truly universal intelligence if we forget to include all of humanity.
