For me, the most serious issue is uncontrolled power. When I say uncontrolled, there are two aspects. There's technological control. We are building AIs that will become increasingly autonomous and won't obey our instructions. And there's governance control, in a political sense. The more intelligent it is, the more easily it can ultimately be used as a weapon against us, against our children, against our democracies. These are all quite serious things that we are not fully grasping.
Yoshua Bengio, we're coming out of a global AI summit organized in India, and more and more voices are raising the question of controlling this tool, while everything is moving faster and faster, with every player announcing new functionalities. ChatGPT, Gemini, Claude, is it going too fast?
It's clear, that's one of the conclusions of the international report I led on AI safety. So there are advances in both regulation and risk mitigation technology, but these advances are insufficient compared to the advances in the intellectual capabilities of AIs. And so the gap is widening. Governments need to understand what is happening.
You say we need to slow down to decide together which applications are beneficial and to target AI research on what we want for humanity, on what will be good for humanity.
Exactly. And there's the aspect of directly beneficial applications, like in health, the environment, and education. And ensuring that AI doesn't cause harm, because today there is an undeniable scientific fact: the recipe used by all the main AI companies for the most powerful AIs doesn't allow us to ensure they will behave well, that they will listen to our instructions. And in fact, that's how we're starting to see all sorts of worrying behaviors in these systems.
We will dissect all the risks you're raising about artificial intelligence. But it's important to start with the distinction you make, because you have a worried discourse on AI, but a month ago, we had in this studio doctors, researchers in the fight against cancer, who explained to us how much artificial intelligence was allowing, and would allow, to detect earlier and treat better. They don't say it's going too fast. On the contrary, they would like it to go even faster.
Of course, of course. In fact, that's part of the research we do in my lab too. The problem isn't how AI is deployed in medicine. Well, there are issues, but for me, the most serious issue is the uncontrolled power that is developing through the intellectual capabilities of AIs, which are measured scientifically and are growing, in some cases exponentially. When I say uncontrolled, there are two aspects. There's technological control. We are building AIs that will become increasingly autonomous and won't obey our instructions. And there's governance control, in a political sense. The more intelligent it is, the more easily it can ultimately be used as a weapon against us, against our children, against our democracies. These are all quite serious things that we are not fully grasping.
Arthur Mensch, the boss of the French startup Mistral, was talking about it. He explained here on France Inter that the real risk of future artificial intelligence is that of massive influence on the way people think and on the way they vote. Do you agree with that? Is it the one weighing on information, on disinformation, that will prevail?
Quite probably. These systems are doing much more than what we saw with social media, in the sense that people develop an almost emotional relationship with their AIs, and so the AI will be able to influence them in a subtle way. We have to trust those who control these systems. But, for example, could Elon Musk's AI influence its users for an upcoming vote, for the midterm elections in the United States next year? That's not in the future, it's already clear. You will interact on X and you will see that it's already the case. So yes, I agree that we absolutely must deal with it, and in general, I believe our democratic institutions are in peril.
AI will give enormous power to those who control it, and concentrated power is the opposite of democracy. What is democracy, actually? Not voting. Democracy is the sharing of power in an inclusive, pluralistic way. But if a few people, a few company CEOs, can influence our political opinions in particular, then the system can't function. It creates considerable democratic distortions.
