Perspectives on AI #1 - Intierview with Gaël Varoquaux

This series of perspectives on AI begins today with a researcher who has been “deeply involved” in AI for many years. Gaël Varoquaux has developed highly specialised expertise in machine learning, which has notably resulted in the development of the Scikit-Learn library—widely distributed as open source—and the creation of the spin-off Probabl, which develops a comprehensive suite of open-source tools for data science. He currently divides his time between the SODA project team, which he leads in Saclay, and the start-up Probabl. Gaël answers questions from Olivier Trébucq.

Interview with Gaël Varoquaux

What is AI? How would you define it?

AI is nothing new; Turing was already talking about it over 70 years ago. It represents the frontier of computer science, which is simply changing its technical underpinnings. I would define AI as the result of an approach that seeks to use computers to solve everyday problems. To do this, AI provides a form of intelligence. 

Are there different categories of AI? Is it a genuine field of research or merely a vehicle for technological development?

Historically, AI has been classified into two categories: symbolic and connectionist. To these we can add generative AI, which, beyond AI’s ability to predict and solve problems, enables the generation of text, sound or images. One sometimes wonders what proportion of AI involves fundamental research versus technical development. This distinction – or even opposition – between research and development is, in my view, unproductive, as computer science has always straddled the boundary between the two. Research aims to create knowledge, whilst technological development aims to produce artefacts. Given that AI enables the creation of complex objects, it constitutes a legitimate subject of research warranting in-depth study. In this sense, it could be compared to the aerospace industry, which deals with highly complex technical issues and mobilises considerable capital. 

Beyond the opportunities it offers, what is the main threat that AI poses to society?

This is the concentration of power around a ‘techno-oligarchy’. In the field of defence, with autonomous systems such as AI-controlled drones, the command structure is simplified to the benefit of a few players at the top, encouraging economic and administrative arbitrariness. This phenomenon of concentration, far from being limited to the military sphere, extends to the economic and political spheres, creating risks for the sustainability of democratic regimes.

Are there different approaches to AI across continents?

There are indeed cultural differences in how AI is viewed. In the United States, players in Silicon Valley are pushing to move forward and see AI as offering infinite possibilities, applying Schumpeter’s theory of creative destruction. In Asia, the vision is of a bright future underpinned by a strong social order. This is particularly true in China. In Europe, there are many concerns about preserving social values that could be undermined by AI. These concerns are being addressed through regulation and standards, intended to act as safeguards against potential abuses of AI. Beyond the legislation itself, attention must be paid to its effective implementation in the face of major players capable of circumventing it. At the same time, we must avoid rigidity that could hinder the growth of European players capable of competing with their American or Chinese counterparts. 

What is the view of American academics regarding the European approach?

I don’t sense any hostility, but rather a perception that Europe is making things harder for itself and putting obstacles in its own way. Many American researchers ask themselves, on an individual basis, ethical questions similar to those of their European counterparts, but the institutional approach differs, with the United States banking on the inevitability of AI deployment, whereas Europe tends to procrastinate.

How can we address bias in training data?

I define bias as a misalignment with a specific cultural or social objective. Since the real world contains biases and training data reflects these inequalities (greater representation of the wealthiest, gender stereotypes such as ‘nurse/woman’ or ‘doctor/man’), AI reproduces these patterns. Take the example of child benefit: depending on whether the objective of those commissioning the AI is to detect fraud rather than identify new eligible recipients, the expected results based on the data used will obviously differ, and the AI will have nothing to do with it… These issues of social statistics are not new, and it is wrong to view AI as a novel phenomenon.

Ultimately, between a blessing and a curse, where do you draw the line when it comes to AI?

It’s hard to say, because in almost every area of application, there will be both positives and negatives. Take cybersecurity, for example: AI will be very useful for detecting vulnerabilities in a system. The flip side is that attackers will use the same tools… When it comes to security, what I fear most – and what is discussed less – is that the formidable development machine that AI will become will produce code of mediocre quality, which could compromise the reliability of systems. 

Extra question. If you wanted to ask an expert a question about AI, what would it be and who would you ask?

What: how can we avoid making mistakes in the race towards AI and prevent future disasters?

To whom: Bill Gates, because he is a major player in the tech world and he has a conscience, which is not the case for everyone in the techno-oligarchic sphere…