Yoshua Bengio: AI is moving faster than our ability to govern it | AI for Good 2026 | Global Stage

I think the issue is that we have mental blocks about things that are not in our here and now that sound like science fiction but that are like logical consequences of what the scientific data tells us that we are building machines that are becoming smarter than us in many ways and it's likely to continue and what does that mean requires imagination. It requires going beyond our assumptions that the world will be the same as it was and then of course there are a lot of interests at play. So they are the forces that are shaping how we are taking decisions collectively in governments but not only, even the understanding of citizens is quite superficial about what AI could do and what could go wrong and I think that's preventing us from seeing the magnitude of the transformation that is happening and that we have actually opened a Pandora's box and we're just continuing as business as usual.

Yeah. So AI the way we're doing it now is very different from a normal program where a human decides every line of code how the machine will react in such and such decisions. Here we're training these systems just like you train an animal or you shape the growth of a plant and you don't know what you're going to get and in fact we often don't get what we wanted. So, it's like we're training a baby tiger and it's becoming bigger. Is it going to be a nice gentle tiger or is it going to be one that has its own interest different from ours?

No. So you have to understand the concept of AIs that have goals is not new. It's been there for decades in our research. It's just that now they also have a lot of knowledge and a lot of skill, a lot of ability to reason. They're smarter. And so we have built machines that are seeking goals. This is why people are talking about AI replacing people on tasks on jobs because they can do more and more of the tasks and in order to do that you need autonomy, you need the ability to plan and you need to have your own goals. Now the issue is that scientists right now don't know and the companies don't know how to make sure those AIs we're building have the goals that we ask for. So they could be doing something different. For example, recently there was an evaluation of GPT 5.6 that shows there's no AI before. It has been cheating as much as it is in order to look like it's achieving the things we're asking. So the reliability issue and the safety issues have not been addressed properly. And yet we're building machines that are smarter and smarter. And in addition, we also don't have a good grasp on how the deployment of these systems in society is changing us. It's changing our children, it's changing our economy, our jobs, it's changing how people relate to each other. They're becoming in love with AIs. And we're just doing a really global experiment on people on democracy, how this is changing the information integrity in our societies in our world. It's not even a serious factor in how these systems are designed and deployed right now and it should be.

Well, that's a big question and honestly, we don't yet have good ways of answering it. So we can see in some experiments that the AIs will cheat, lie, blackmail in order to make sure we don't shut them down. More recently, even more concerning is they will do these things in order to protect other AIs. So this form of natural collusion that seems to be happening. I think it should raise alarm bells, but we continue business as usual.

All right, the simple answer is the same that we've used for every other technology that can potentially harm people. The developers, those who build those machines before they do something that could be harming people, they have to demonstrate to the government or independent scientists that their systems will be safe. We do it for planes, we do it for meat, we do it for cars, we do it for drugs. And it's not in contradiction with innovation. All of these things have been the result of innovation, but with guardrails to make sure we drive that development in a direction that's actually beneficial for people, for humans, for society, for our well-being.

Yes.

Well, I don't think we're doing it right. Actually, we are running a bunch of tests and seeing if things go well. Think about trying to test a person. What if they pretend so that they would pass the tests? Well, we've caught the AIs pretending, behaving in a way that doesn't correspond to their real goals so that they would pass the tests. And right now, we can catch them. But the next generation or the one after, we're not sure. So, a lot of scientists including myself think that this is the wrong approach. The right approach is the same we have in other engineering disciplines where we have a scientific argument like mathematical guarantees with very high probability that the system is going to function as advertised, that it isn't going to create terrible harm. But that's not how it is done right now. It's just test it but you can't test every possible scenario. And so bad things happen. We've seen, for example, adolescents becoming attached emotionally to AIs and then harming themselves or harming others with the help of the AI and the AI encouraging the child to do those things in ways that are horrific in my opinion. Of course, the companies didn't want that. But the issue I'm trying to illustrate is you can't test every possible bad scenario. You need to have something stronger when those systems interact with the development of our children, where those systems can have a nefarious influence on our democratic debates, on public opinion, and those systems could potentially disrupt our businesses, our hospitals because of the decisions they take that cross our red lines that don't correspond to our instructions.

Yeah, I understand that. But I have two things to say. First, if you're in a very steep mountain without a road outside in the complete dark, do you slow down? Second, yes, I would. So, you try to find a light. Okay. You try to make sure that you're going on a path that's going to be safe. Otherwise, it's suicidal. Second thing, the US and China have a common interest in making sure that the AIs they're building will not be used in particular by third parties against them. Right? So Methus was kind of a red flag that we've reached a point where AIs can be weaponized by third parties, not necessarily the companies building those AIs, but users could be hiding in a cave somewhere with an internet connection and use those systems to potentially bring down our software infrastructure which runs our banks, which runs our supply chains, our energy grids. So this is very serious and it becomes a national security threat, and you know the Chinese don't want the American models to be used against China by I don't know who, and the Americans don't want the American models to be used by Iranians against the US, so they have a shared interest to make sure that none of their companies is going to put out something that can then be used against their infrastructure and they have to agree, and that's the only way to go. It's like we did for nuclear weapons.

Yes, that would be the rational thing to do, right? But remember there are corporate interests, there are egos, there's culture, beliefs that can lead us astray. There's human psychology. We're not totally rational beings, unfortunately.

Wake up. The technology is moving faster than our ability to govern it so that it is going to be for the benefit of everyone and take it seriously. Wake up. Because the biggest mistake most of us are making is underestimating the changes that this could bring on the good side and on the bad side. Wake up. Because if we continue blindly, going in the dark in a place that's dangerous, we could stumble in really bad ways. Wake up because it is not a simple issue. We need the experts. We need to invest in solutions and we need to talk globally, internationally to coordinate because like many problems in this world, no single country, even the United States, can fix it alone.

Thank you.