Laurence Devillers

Home automation, pet robots, connected speakers... Artificial intelligence is part of our daily lives. But this digital revolution brings its own set of ethical questions. For 30 years, Laurence Devillers has been one of the most visionary actors and critics.

Until 20 years ago, few consumer applications were based on artificial intelligence (AI). The rise of GAFAM¹ at the end of the 20th century has reshuffled the deck by anchoring it in our habits, sometimes to our detriment. While it has allowed great advances in research, AI raises many ethical questions. Laurence Devillers, a professor at LISN², is part of the generation of researchers who accompanied the beginnings of AI. She believes that this technology can teach us a lot about the birth of language.

From computer science...

Born in Burgundy in 1962, Laurence Devillers is steeped in science. Her mother is an associate professor in plant physiology at the University of Paris3 and her father, a centralist with a strong interest in ethics, works at the CEA4 and the IPSN5. From a very young age, she asked them questions about the human being and consciousness.

She obtained her baccalaureate in mathematics and physics in 1981 at the Lycée Lakanal in Sceaux, where neuroscience was just developing. She went on to study for her first MIASS6 degrees, one of the courses of which was linguistics, and computer science at the Sorbonne. "Computer science seemed very open and dynamic, and there were almost as many girls as boys," recalls the researcher whose parents did not want to make any difference in their children's education.

She could already see herself designing objects and places by computer simulation. But in her DEA⁷ at the University of Paris Saclay, she took a course that would seal her fate as a researcher. It dealt with phonetics, interpretation of voice spectrograms and other aspects of language and its modeling. "It was extraordinary! From there, I moved on to researching language using automatic processing tools." The teacher of this module, Jean-Sylvain Liénard, then director of LIMSI2, would be her thesis director.

...to language

Laurence Devillers began her thesis on European projects dedicated to acoustics and the first Markov8 learning models applied to linguistics. "It was not a personal choice, I had to wait until I found an innovative field of research that would interest me. In this laboratory, linguists and computer scientists cohabit. "At the time, it was not enough to teach language to the machine in order to model speech recognition; we also had to understand the different levels embedded in language." This was the emergence of neural networks, revolutionary systems based on an artificial neuron that mimics the nerve cell and its connections. This fascinated her. Especially since the hybridization of the two techniques enabled a more refined phonetic recognition. This was the subject of her thesis⁹ (see box), defended in 1992. "With these biomimetic approaches, we were building machines that recognize what we say without any linguistic knowledge. Even if the performance of these systems is close to that of humans, they are meaningless for the machine."

A committed researcher

After a stint at the Paris branch of the Stanford Research Institute, where she worked on a French language learning system, she was awarded research contracts and then a position as an associate professor at the University of Paris-Saclay in 1995, where she developed human-machine dialogue systems, notably for the SNCF. In 2000, she studied the modeling of emotions in the voice and, in 2006, obtained her habilitation to direct research on this subject.

In 2012, after a year at the Sociology Department of Sorbonne University, she joined the LISN where she leads the Affective and Social Dimensions in Spoken Interactions team. "I have always navigated in areas where we subtly handle social interaction, psychology and cognitive behavior to improve voice and emotion recognition systems. It forces you to be curious and that's exactly why you do research."

In 2016, she took it to the next level and looked at what systems incorporating AI do with the emotions they detect. Ethics is her new hobbyhorse for dealing with the challenges and ubiquity of digital technology in the early 21st century. "These systems are incredibly beneficial for people who are depressed or losing their autonomy, but I realize the strong manipulative power of these objects that call you by your first name and claim to love you. Profiling people based on their emotions and behaviors is also being able to influence them. We need safeguards so that those who create these systems don't go too far." A small nod to GAFAM, the primary objective of AI is to serve health and education, among other things. "I'm not looking to regulate but to lay the foundations for good practices in AI".

From 2017, she joined the Allistène¹⁰ ethics committee and then the National Pilot Committee for Digital and AI Ethics, responding to missions from the Prime Minister on the ethics of chatbots¹¹ in particular, participating in the Global Partnership on AI (GPAI) on the future of work, working with AFNOR¹², speaking at public conferences, chairing the Blaise Pascal Foundation for mediation in mathematics and computer science, and getting involved in the inclusion of women in science. For while she warns about dependence on machines, she also insists on the fact that "80% of coders and AI designers are men and 80% of the systems they design have female voices, names and bodies". Commitments that she says earned her the Legion of Honor in 2020.