EU Leading on AI Regulation, Says Canadian Computer Scientist Yoshua Bengio | Global AI Lens

Well, the problem is there are many threats. You have to understand that intelligence gives power, and the question we don't know how to answer properly is how is that power going to be used, by whom, and is it going to be used according to our interest? Is AI going to behave according to the instructions we give it? Unfortunately, we don't have the good technical answers right now. We in fact see in the lab experiments where the AI will be deceptive. They will be willing to blackmail an engineer so that they will not be shut down and replaced by a new version. They will try to pretend that they don't know something so that we will keep them. So they know that they're being tested and they behave in the ways that we want to hear, not necessarily the reality. So there's lots of issues. Companies don't know how to prevent these AIs from being used by criminals for cyber attacks. There's been a number of waves of these attacks. There's concerns that they could be used by non-experts to create bioweapons. So these are really, really serious. And we need to figure out how to organize society and also change technology to deal with these challenges right now.

There is very little regulation globally on AI. We are starting to see some movement. The EU is probably leading on this. There's a couple of states in the US that are moving forward to some extent with transparency-based regulations, and the EU is also mostly a transparency-based regulation. We need all the companies to tell the world, and especially the government, what is the risk management process that they're following in order to evaluate the risks that their systems can create in society. And then, of course, if those risks are not acceptable, what they do to mitigate those risks, as we would ask for almost any technology. You know, if you're using a drug, if you're taking the train, there are regulations to make sure the companies that are building these things are doing it in a way that's not going to be dangerous for the public.

So we think that the current way AI is being trained leads to these problems naturally because in the first phase of training, the AI is trained to imitate what people would do in the same circumstances, which means they incorporate a lot of human drives, which may sound fine except that we don't want to build tools that don't want to be shut down, right? For example, or tools that are going to be lying to us for some reason or another. So what we are working on is how we can train these systems so that they will be completely honest, that they will not claim something even though they know it's not true, and then when they're not sure, they would say, 'I don't know.' So this would be a big advance because we could then ask the following question to the AI automatically for every action that it wants to do: Is this action going to harm people or is it going to violate one of our norms, one of our laws? And if the answer is yes, then the action is rejected. So we could build systems that are safe, which is something that we need to deploy AI in a way that will be good for people. We need technology to be reliable. We don't want flaky technology that, you know, we could lose control of or could be used to attack other people. And so on. So it's urgent to figure out these technical questions.