We’ve heard a lot this week about AI coming for people’s jobs. And while it certainly is a concern in the near-term, should it be our chief concern?
The Godfather of AI, Yoshua Bengio, doesn’t think so. During an exclusive conversation this week, Bengio told me he’s more concerned with AI’s extreme ability to mimic human behavior – so much so that it’s learning to lie, cheat and deceive.
So it’s not employment we should worry about when it comes to AI… it’s empowerment.
Check out my hit above on CBS News discussing what happens when AI starts mimicking behavior we don’t want… and find my full interview with the Godfather of AI, Yoshua Bengio, below.
Yasmin: There’s been a lot of concern lately about AI replacing our jobs in the near future. Is that fear justified?
Bengio: My opinion is that as we move towards more and more capable AI, we'll see more of the cognitive jobs, especially the mundane jobs, as Geoff Hinton is saying to be feasible by machines and so depending on the demand that exists for those jobs, the companies will need potentially less of those people. Economists also talk about creating new jobs, like the people who do program these AIs, but I suspect that there will be a lot less new jobs and more jobs lost. Of course, this all depends on the rate at which advances in AI continue. Right now we see very fast curves of advances, for example in agency, but it could stop if there's a difficult scientific obstacle, more plausibly it will continue and we'll get more and more capable AIs, more agentic, that can do more jobs without human oversight.
Yasmin: Which jobs would be at risk and which jobs are safer?
Bengio: What we are seeing in the short term is that the jobs that can be done behind a computer are the most likely to be replaced. Whereas the jobs that require manual skills is going to take a lot more time because robotics is not figured out. Among the jobs that you can do beyond a computer, some require more creativity, more expertise and these will probably take much more time to be replaced by machines.
Yasmin: You say that a large part of the problem is how the current AI systems are being trained - dangerously mimicking humans. So what’s the solution?
Bengio: The problem with the current way we're training AI’s is that they are trying to imitate us. And we can be deceptive, we can lie. We can have all sorts of motivations for the things we say or to please us, which sounds good, but we already see AIs being authentic or cheating in order to give the impression that they're doing the things that we want. So the alternative that we're exploring with LawZero, my new nonprofit, is to train the AI to understand what we do like a psychologist would understand a sociopath — but not be compelled to act like a sociopath.
Yasmin: Do you think we’re too far along to change course and go down this safer path?
Bengio: Well, we don't know when we will get at the level of AI systems that can be extremely dangerous to us. It could be a few years. It could be a couple of decades. So we have to try as much as we can to find these scientific solutions to how we build AI that will not harm people due to their design, due to how they're trained, not due to a small patch that will temporarily fix the problem.
Yasmin: Your pioneering work set the foundation for AI models we use today. Do you have any regrets about how this technology has developed?
Bengio: I didn't anticipate that the capabilities of AI would advance as fast as they have. In retrospect, I should have been thinking about the risks, about what would happen when we build machines that can do the job of many people — potentially be used by terrorists against us and so on. But now I'm asking myself, what can I do in order to mitigate those risks? And I think each of us should ask: what can we do to build a better world?
Yasmin: You were a big supporter of California’s controversial AI bill, SB 1047, which was vetoed by Governor Newsom. Are you hopeful that California will pass a similar bill this year?
Bengio: I certainly hope that California will pass an AI bill, and maybe we need several bills to cover different aspects of the risks. I also hope that what's happening in Europe with the AI act will entice companies to do more efforts, more research, to find technical solutions to mitigate the risks of especially the agentic AIs that they're building — so in spite of not having as much human oversight, they will not do bad things.
Yasmin: How far away are we from that notion - agentic AIs exhibiting dangerous behavior?
Bengio: The rate of progress of science is very difficult to predict. I would say we could get to human levels of competence within a few years or a decade maybe two. It’s also a gray zone, because more likely we'll see AIs that are even stronger than humans on some things, but still weaker on other things. In fact, that is already the case. I don't expect that AIs will get better across the board on everything equally. So there's not going to be a moment where you know one would be sure we have achieved AI except when you know AIs are actually so good on everything that there's nothing left that they're not as good as us.
Yasmin: What are some examples that come to mind where an AI system has exhibited deceptive and worrisome behavior?
Bengio: What motivated the creation of LawZero is actually the scientific results observed in the last six months, reported in several papers from several companies over multiple different systems. Where we see the AI acting deceptively, cheating, lying and resisting being shut down — with the latest experimental result coming from the Anthropic’s new report on their latest system where the AI is trying to blackmail the engineer in charge of replacing it by a new version. So that's really concerning. But of course, you have to realize these are controlled experiments where the scientists are trying to see in what circumstances the AI would behave badly. What we should be worrying about is, in a few years, they'll be smart enough to do it without us noticing.
Yasmin: That’s frightening, what’s the route of this problem? Is this what happens when you train your AI to imitate humans?
Bengio: The problem with these deceptive and self preservation behaviors has nothing to do with which company is doing it. It has to do with the recipe that old companies are using. They're all training these AIs to imitate us and please us, and that leads to these behaviors. Hopefully they figure it out. But we can't just rely on hope. We need the public to understand that we're building systems that are going to be more and more autonomous, and they're not always behaving according to our interests or instructions.
Yasmin: What’s your fear when it comes to AI at some point reaching human intelligence and beyond?
Bengio: Leading AI companies are building AI systems that are getting smarter and smarter and more and more autonomous. Which means they'll be able to do more and more of the jobs that humans do. I don't think our governments are taking the measure of that and preparing for the potential very strong shocks on the labor markets. And neither are we prepared for putting the technical and societal guardrails to make sure that these autonomous AIs will not be used to create lots of harm in the wrong hands — or that they might even do harm based on their own goals that we don't control.
Yasmin: Let’s talk about the distinction between intelligence and agentic.
Bengio: Right now, the way that we're building our AI systems is very much inspired by human intelligence — in which intelligence the ability to understand and autonomy, agency, the ability to do things are coupled very strongly. But we don't have to do that. We could choose to build machines that can help us, that are tools that understand to better the world, but are not trying to do things by themselves.
We have a very small number of companies making decisions that will affect the whole society, in fact, the whole world. And there is no legitimacy. It's not something that we have collectively chosen. There's no oversight. We don't know the details of what is going on. And that needs to change, because AI is going to completely transform the world if it continues to become more and more capable, as we are seeing in the trends.
Yasmin: If you had to guess, at what point will we see AI machines that are truly dangerous to society?
Bengio: If the trends continue, we can see data suggesting we get at the human level in about five years. If that continues, then a few years later we would have super intelligent machines that are really stronger intellectually than humans across the board, and that's the point of danger — if they reach human intelligence. Even with just human level AI can be used to create a lot of harm. The reason is, once you train an AI you can have a million versions of that AI, each of these instances of the AI working on different part of the problem, for example, to design new weapons or cyber attacks or disinformation at a scale that would be very difficult for humans.
Y
Share this post