Godfather of AI warns humanity: It’s all hurtling out of control

Geoffrey Hinton, the British-born “godfather of AI”, has warned that the technology is “hurtling out of control” and “humanity does not know what to do to make it safe”. The Nobel Prize-winning computer scientist said a disorderly global race for AI dominance is putting the “future of humanity... on the line”.

He said it was perilous to let a handful of US companies dominate: “It’s particularly dangerous because of the characters of some of the people, like [Sam] Altman and [Elon] Musk and [Mark] Zuckerberg. I wouldn’t buy a used car from any of them.” Hinton, 78, quit a role at Google in 2023, in order to speak about the dangers of AI.

He invented the concept of the neural network – the system underpinning AI models. Hinton was in London to deliver one of his starkest warnings yet – ahead of the release of a documentary, AI: Probably Nothing To Worry About . The film airs on Channel 4 next Thursday, 12 November, at 10pm.

Here are his candid, arresting views on the future of AI – both the dangers and opportunities. When you look at the exponential growth of AI – and how it’s suddenly racing towards superintelligence – do you think AIs might learn from the mistakes of humanity? Or are they doomed because they are built on the win-at-any-cost mentality of their creators?

That depends a lot on how we create and how we train it. At present, for example, we train it to predict the next word in any document you could find on the web. That includes, presumably, the diaries of serial killers.

It will be a good idea to train it just on things that exhibit what we regard as good behaviour, and not showing any bad behaviour until it knows what good behaviour is – and knows when it sees bad behaviour, “that’s bad behaviour”. If you’re raising kids, that’s what you would attempt to do.

It is not always possible. But you’d attempt to show them good behaviour, and then later on, tell them about bad behaviour and tell them it’s bad. But showing them this mix of everything, and training AIs so they have to be able to adopt the personality of anybody who wrote a document, is a crazy way to train a being.

Geoffrey Hinton speaks at an event in London ahead of the documentary AI: Probably Nothing to Worry About (Photo: Clare Hannan/Hannan Images) You do get the impression that it’s all hurtling out of control. Is that a fair impression? I think it’s probably because it is all hurtling out of control.

Can governments, or the corporations that are building superintelligence, control its development, ultimately? We don’t really know. Everything’s extremely uncertain.

What’s happening at present is there’s this competition between countries, and between companies, to make it smarter and smarter and smarter, in the hope that you can replace lots of workers and get lots of money. That’s what’s driving it. And this is the future of humanity that’s on the line.

Not all of the experts, but nearly all the experts, believe it’s going to get more intelligent than us. They disagree on when. Some say next year, some say five years, some say 10 years.

I used to say 10 to 20, now I say 10. It’s always getting closer. And what we ought to be doing is saying, okay, we would like to [still] be around.

We’re going to have to share the planet with something smarter than us. What would we like the other aspects of that being to be, other than its intelligence? And we hardly do any work on that.

It’s crazy. Meta CEO Mark Zuckerberg was among the tech bosses criticised by Hinton (Photo: Carlos Barria/Reuters) With every country fighting for sovereignty in the tech space, how dangerous is it that we rely on these five companies from the US? It’s very dangerous.

And I think it’s particularly dangerous because of the characters of some of the people involved, like Altman and Musk and Zuckerberg. I wouldn’t buy a used car from any of them. Are you to blame for any of this?

For the dangers that have been created? I should have thought about safety sooner. I always believed it was way in the future.

What do you feel about the comparisons with Oppenheimer? (American physicist J Robert Oppenheimer , who led the development of the first nuclear weapons. Hinton has previously rejected the comparison, noting that thousands of other scientists also work on developing neural networks.) A week after it was announced that I had won the Nobel Prize, in a subject I don’t do, namely physics, I was giving a lecture to a bank in New York to make money, because I’m unemployed at present.

And the last question, after I had given the lecture was: “Later in his life, Oppenheimer came to regret what he’d done. How would you compare yourself to Oppenheimer?” And I couldn’t resist, suddenly it flashed through my mind, and I said: “Oppenheimer never won the Nobel Prize in Physics.” Geoffrey Hinton receives the 2024 Nobel Prize in Physics from King Carl XVI Gustaf of Sweden in Stockholm (Photo: Henrik MontgomeryTT/AFP) How likely is a “kill switch” scenario? (An emergency shutdown mechanism, to prevent loss of human control over AIs.) I don’t think there is a kill switch scenario.

When AI is smart, it will be very good at persuasion. So even if you have someone who’s meant to turn it off if it misbehaves, it will be able to persuade them not to. We have seen examples of people – without performing any physical actions themselves – getting things done just by talking.

My favourite example is January 6, when Trump managed to invade the Capitol without going there, just by talking. Imagine an AI that’s really smart, and can convince people of things. And you’re in charge of the switch, and the AI talks to you, and explains to you why it will be a really bad idea to turn it off, and you’ll believe it.

So I don’t think a big switch is really an option. Do you believe that artificial intelligences should enjoy civil liberties? No.

In 2023, I realised these digital intelligences we’re developing are actually a much better form of intelligence. They can share much better than us. Not just a little bit better, but hugely better.

If you have many copies of the same model, they can share what they’ve learned at trillions of bits per second, and we can share it at about two or three bits per second, maybe 10 bits a second. (Which is what I’m doing now, talking.) So they are a better form of intelligence. And I thought at that point, hey, okay, so we [humans] were a passing phase in the evolution of intelligence.

We’ve now produced these things. They’re going to take over. They’re a much better form of intelligence.

Because we’ve been the apex intelligence on the planet, we’ve been able to live with the idea that we should have more rights than other animals, because we’re smarter. That was fine when we were smarter. We might want to revise that opinion – that rights go with being smart.

I now think: we’re humans, we care about humans. What we should do is try and create these things so that they’re very good for us. So they care more about us than about themselves, and they have less rights than us.

So I’m willing to give them less rights, even though they’re beings. I’m willing to be humanist. Doctors use a surgical robot to perform a knee joint replacement in Qinhuangdao, Hebei Province, China (Photo: Cao Jianxiong/VCG via Getty) You say that on a bad day, you think about ways that AI could destroy the world.

And on a good day, you think about how it can solve illnesses. How else might AI solve some of the world’s biggest problems? And what do you think is needed to make that happen?

AI can invent new materials. It’s already been used to invent interesting new materials. So, for example, if AI could invent much better batteries, then we could get most of our power from solar and wind.

Solar gives plenty of power. The problem is it’s intermittent. Well, maybe not in Manchester, but anywhere else...

If we could store it, that’s a great way to have power. You wouldn’t actually need nuclear fusion, then, if we could store solar energy. The way to build a fusion reactor is – because they get very hot – it’s to have it about 90 million miles away, and have it refuel itself.

And just collect the energy from it. And if we could design materials for better batteries, that would be a big win. So that’s one of the big wins which might help with climate change.

It might even offset the energy all these data centres are using. That’s just one example. I think, almost everywhere, AI can make us more productive.

But the question is how we handle that. This increase in productivity should be able to provide a huge increase in wealth. But Sam Altman and Elon Musk talk about radical abundance from all of this.

They don’t talk about exactly how the wealth is going to be distributed. We have a pretty good idea what they think is going to happen. And they all seem to believe in tech feudalism.

Certainly Musk does, which is why he’s so obsessed with Mars. They believe we should get rid of nation states, and we should have states run by CEOs who make all the rules. And they are Übermenschen [supermen] and ... everybody else works for them.

Do AIs have a sense of morality? And if so, does this vary between different AIs? If you train them on data that exhibits good behaviour, that’s what they will understand to be the right way to behave.

And in that sense, they’ll have a sense of morality. My belief about morality is not that it came from God, but it came from groups of people interacting. And if you’re going to interact within a group of people frequently, it pays to tell the truth most of the time, and it pays to honour your word, and so on.

And if you look at how we behave, what you see is, when we’re interacting with people who we’re never going to see again, and the stakes are high, we tend to be rather immoral. When people are selling houses, they’re perfectly happy not to actually reveal all the little facts about the house that weren’t told to them when they bought it, and they’re not going to tell the next person.

Because they’re only going to have to deal with that person once. They’re never going to see him again. But it’s not what you would do if you were selling a house to a friend, you would probably tell them what was going on.

So, our morality is very much a functional thing, and we have evolved it so that we can co-operate effectively. The SkyFall drone company’s StarGate drone in Ukraine. It has AI features and is used to rescue injured soldiers (Photo: Evgeniy Maloletka/AP) Which aspects of human intelligence would you say AIs are less good at, currently?

And what would it take for them to make that breakthrough? Spatial intelligence. So...

I have a house with a big lot and a big back yard, and in Toronto you can build a garden suite in the backyard, if you’ve got enough space. So quite recently I asked ChatGPT 5.5: “How long does my backyard need to be in order to build a garden suite?” And he came back and said: “You need five feet between the garden suite and the back fence, and you need 19 feet between your house and the garden suite.

So your back yard needs to be 24 feet long. Which means the garden suite is exactly zero feet thick.” So that’s the kind of intelligence that it’s not good at. And if we train it on more spatial things, it’ll get much better at that.

People like Yann LeCun [Meta’s chief AI scientist] say we need world models. [World models use knowledge of physics, spatial awareness and cause-and-effect to predict the next physical change to an environment.] That’s to get at that spatial aspect of it, it’s still weak there. What’s amazing is that [the current] models trained just on language [predicting the next word in text – LLMs or large language models] can actually learn quite a lot about space.

But you learn much more about space if you can pick things up and move them around, and you can walk around, and you can see. So that’s an aspect where we are still better than AI. Scientist Timnit Gebru has warned about the danger that AI can intensify bias and discrimination (Photo: Kimberly White/Getty for TechCrunch) To what extent do you feel that biases and discrimination form part of the knowledge base, and therefore inform some of the moral battles that you’re talking about – the difference between right and wrong?

I’m an old white man, so I’m not particularly sensitive to bias and discrimination, because they all go in my favour. Obviously, if you train an AI system on data where people are exhibiting bias, the AI system will have the same bias. There is something you can do that’s quite positive, which is, with an AI system, you can “freeze the weights”, and then you can examine what its bias is, and you can try to reduce that bias.

You can’t do that with people. With people, once they figure out that you’re figuring out what their bias is, they change their behaviour. So I’m not as worried about bias and discrimination as other people. [Computer scientist] Timnit Gebru, for example, who famously left Google [and warned about the biases and ethical risk in LLMs], she’s very worried about that.

She has better reasons than me to be worried about it. But I am an old white male. How could we tell the difference between something that perfectly mimics having a mind or having experience, and something that actually has a mind or experience?

The crucial word there is experience, and I think almost everybody has a completely wrong [mental] model of what “experience” is. So... almost everybody has a model in which there’s a theatre called “my mind”, and in this theatre called my mind are things which I experienced directly.

So when I see an apple, there’s an image on my retina that’s upside down, and then there’s various goings-on in my brain, and then somehow, in this inner theatre, there’s the apple I actually experienced, and that’s where I have the experience of the apple. I think that’s just a model.

Everybody is so convinced of it, they think that’s manifest truth. It’s just a theory and it’s a very wrong theory. It is as wrong as the theory that the Earth was made 6,000 years ago.

I think: when I see an apple, what I experience is the apple, the real physical apple is what I experience. I experience that apple by having an image on my retina and various kinds of activity in my brain, and that’s what goes on in order to “experience” that apple. But there’s nothing inside.

Experiencing the apple, there’s these patterns of activity inside, but they’re not like photographs. So I think people treat the phrase, “I have a subjective experience of something”, as if it was the phrase, “I have a photograph of something”. But if I have a photograph of something, I can ask: “Where’s the photograph, and what’s it made of?” And people think: “If I have an experience of something, I can ask, ‘Where’s the experience, and what’s the experience made of?'” The philosopher will tell you, well, the experience is something in your mind.

It’s an inner thing that only you can see, and it’s made of qualia . [Instances of subjective experience.] I think that’s all just rubbish. I don’t think we have inner experience. I think we experience things in the world.

We do that by things going on in our brains. A female humanoid robot wearing intelligent bionic skin at a Unitree Partner Ecosystem conference in Hangzhou, China. Hinton does not believe AI should have rights like humans (Photo: Long Wei/VCG via Getty) Suppose I take some LSD, and I tell you I’ve got the subjective experience of little pink elephants floating in front of me.

Some philosophers would take that to mean: there’s an inner theatre called my mind, and in this inner theatre, there’s little pink elephants, and if you ask: “Well, what are they made of?” the philosophers are saying: “They’re made of qualia. They’re made of pink qualia, and elephant qualia, and floating qualia, and not-that-big qualia, and trunks-down qualia, and all-facing-to-the-right qualia, all stuck together with qualia glue.” That’s not my theory, as you can probably tell.

Here’s a much simpler explanation of what I’m saying, when I say I have the subjective experience that there are little pink elephants floating in front of me. What I’m saying is – and I’ll now tell you the same thing, without using the word subjective experience... “ I believe my perceptual system is lying to me , but if it wasn’t lying to me, there would be little pink elephants floating in front of me.” That’s the same thing.

And notice, there’s no internal pink elephants there. It’s just hypothetical, external pink elephants. What’s funny about subjective experiences, is it’s hypothetical.

Like when I imagine something. The thing isn’t really there. It’s hypothetical.

But it’s not in my inner theatre called my mind. When I imagine something, what I imagine is something out there in the world that’s not actually there. And I will now show you a chatbot having a subjective experience...

So, I train up a chatbot. It can point to things, it can talk, it can see. I put an object in front of it and say: “Point to the object,” it points to the object.

No problem. I then put a prism in front of its lens. And I put an object in front of it and I say: “Point to the object,” and it goes like this. [Points in the wrong direction, off to the side.] And I say: “No!

That’s not where the object is, it’s actually straight in front of you, but I’ve put a prism in front of your lens.” And the chatbot says: “Oh, I see. The object’s actually there. [Points in the correct direction.] But I had the subjective experience that it was there [wrong place].” If the chatbot says that, it’s using the words “subjective experience” exactly like we use them.

And so, when you have a subjective experience, it’s not what you think you have. It’s not this inner thing . And chatbots don’t have this inner thing either.

But they have subjective experience in just the sense that we have it, which is not the sense you believe it. A humanoid Unitree robot plays ping-pong at the World Robot Conference in Beijing (Photo: Kevin Frayer/Getty) The co-founder of DeepMind, Demis Hassabis, has a very competitive nature.

Is it true that you and Demis had a game of table tennis, and that you won? Yes. I visited DeepMind back in 2015, and they had a ping-pong table there, and people were playing ping-pong, and I love playing ping-pong, and I’m quite good at it.

Not very good, but quite good. And I said: “Does Demis ever play?” And they said: “No, Demis never plays, he only plays games that he wins.” And so I managed to convince Demis: “You know, Demis, I’m a lot older than you, and I’m okay at table tennis, but I’m not much good. Should we have a game?” And he said yes.

And we had a game, and I won, and he really, really, really didn’t like it. Do you think that there’s an argument for making either the people who invented an AI, or the AI itself (in some fashion), legally liable? Yes.

Obviously, if people produce AIs and they do bad things, somebody should be held to account. I think it should probably be the people who develop the AI. There should be legal liability, yes.

Did you think it was a good idea that Sam Altman was removed? Did you have any influence in that? (The OpenAI co-founder was fired in a boardroom coup in 2023, before being reinstated after an investor backlash. Hinton’s former protégé Ilya Sutskever, OpenAI’s other co-founder, led the coup.) No, I had no influence.

If Ilya had only called me, I would have told him: don’t do it about a week before they’re going to have a big new funding round, and nearly all of the 700 employees are going to be able to turn paper money into real money. They were going to be allowed to sell stock. So nearly all the employees were about to become millionaires, and that’s the moment Ilya chose to fire the chairman.

Not a good idea. Ilya has a strong moral compass. He’s a brilliant scientist.

He’s not so strong on political manipulation. Sam is the exact opposite. So would you remove Sam Altman if you had the choice?

Yes. But after the funding round. The decisions of Open AI CEO Sam Altman could be key for the future development of artificial intelligence (Photo: Justin Sullivan/Getty) To what extent do you think there’s a kind of determinism at work here?

Is all of this inevitable? Is this just how society’s going to evolve? I have no idea.

But my view is: right now is a very important time in history, and we should all try to do what we can to make it come out the way we would like it to come out. Just because you think it’s going to be determined historically one way or the other – I don’t think that you should not act.

The UK hosts the G20 next year and Andy Burnham has made AI the central theme of this world leaders’ meeting. So what should the UK do with its G20 to help on this issue? I’m not a policy person, but the UK has tried to position itself as very important in AI.

I don’t think it’s going to be.