Nvidia CEO Jensen Huang has a simple analogy for how to make artificial intelligence safe. He says that people will have to create new social norms and adapt society to the technology, the way cities developed sidewalks and crosswalks to protect children who once played in the streets.
This ignores the fact that traffic deaths remain the leading cause of death for those aged five to 29, killing about 180,000 annually worldwide. But to Geoffrey Hinton, a professor emeritus at the University of Toronto, there’s a more fundamental flaw with the argument.
Professor Hinton, the Canadian “godfather of AI” who split from Google so he could speak freely about the technology, says the difference is that “cars aren’t going to gang up on you and take over ... AI is just far more powerful.”
Geoffrey Hinton, known as the godfather of AI, visited The Globe and Mail editorial board in July to discuss the promise and peril of artificial intelligence.Fred Lum/The Globe and Mail
Still, he has his own automotive analogy, one that doesn’t involve contorting society to fit AI. For him, safety starts with AI regulation , and that doing so should not be perceived as a brake on the technology but as a steering wheel.
“The whole point of regulation is not to prevent you innovating but to make sure that you innovate in a direction that makes the technology help people rather than hurt them,” he says.
Earlier this month, Prof. Hinton visited The Globe and Mail editorial board for a wide-ranging discussion on the promise and peril of artificial intelligence, a technology rapidly becoming more capable and on course to become smarter than humans.
And to be clear, there are advantages to AI as it exists now. There are medical breakthroughs, such as new drugs to treat Parkinson’s and rare diseases. There are productivity gains, as many individual users can attest.
But the benefits can’t distract from the potential dangers. While the day when AI outstrips the human intellect is still unknown, the estimated timeline to that moment is rapidly shortening. And what happens after that is hard to predict.
Prof. Hinton won’t give his own estimate for the likelihood of an apocalyptic or extinction-type scenario. But he says that, based on a range of opinions, the best expert guess is 10 or 20 per cent.
Which could make AI one of the most dangerous technologies ever invented. It is far too risky to leave its development and rollout in the hands of the techno-libertarians of Silicon Valley. Even short of a cataclysmic future, AI will lead to employment and social upheavals. And it could also lead to a crisis of self, forcing humanity to confront how to manage – and relate to – what it has created: a new form of being, neither human nor animal.
Prime Minister Mark Carney announces the new federal AI strategy in Toronto on June 4. The "AI for All" plan calls for widespread adoption and education about AI.Cole Burston/Reuters
The federal Liberal government believes such concerns can be brushed aside. The recently released AI strategy assumes that benefits of the technology outweigh the risks. The plan – dubbed AI for All – calls for widespread adoption.
But many Canadians remain unconvinced. This is a technology that only a minority of people use regularly but about which a majority have concerns.
An international survey by Ipsos showed that Canadians were the least likely of the 32 countries to say that products or services using AI have more benefits than drawbacks. And they led the table among countries most likely to say that products or services using AI made them nervous. In both cases, the negative feelings are deepening, compared to surveys done in previous years.
To the Liberals, this is an indication that Canadians don’t understand the technology. That reaction is patronizing. Canadians are legitimately concerned. They may not fully grasp AI, but that doesn’t mean the fears are unwarranted.
The Canadian government’s desire to convince everyone that AI is good and useful and inevitable is also at odds with a growing public movement in the other direction. The backlash against data centres is one example of this. While AI itself is amorphous and hard to target, data centres are its physical manifestation and they can be protested.
Companies wanting to build data centres, and municipal leaders ready to accept them, must understand the fears. A better sales job is possible. Centres can be built to recycle water and direct waste heat to nearby homes. But people need to be convinced. Simply telling the public that this future is inevitable does not make that case.
People march in protest against the opening of AI data centres in Vancouver in June.DARRYL DYCK/The Canadian Press
That’s in part because the future being presented by AI boosters themselves is scary. If these cheerleaders are correct in their predictions, the technology will lead to massive job losses. Entire types of employment will disappear, including many forms of entry-level white collar work.
How will people earn a living in this future? Where will governments find tax revenue? How will people find the meaning and purpose that a job can provide? It’s unclear.
While former technological breakthroughs, such at the Industrial Revolution, created new types of jobs, that process could take decades and the intervening period was often grim.
“I think you should read a bit of Dickens before you just blandly say it turned out well,” Mr. Hinton said, going on to quote William Blake: “They were dark, satanic mills.”
With the stakes this high, the people trying to push this future cannot be the ones solely in charge of how it develops. Self-regulation is never a good idea. It’s a less good idea with a potentially devastating technology. And it’s still less of a good idea when one considers the mindset of some of the top people in Silicon Valley.
Elon Musk, whose chatbot Grok was embroiled in controversy over its ability to generate faked nude images of people without their consent, is on the record as saying that “the empathy response” is a “fundamental weakness of Western civilization.” Meta head Mark Zuckerberg imagines a future in which most of a person’s friends are actually AI creations, a dispiriting scenario that he thinks can solve the loneliness crisis.
In response to criticism, OpenAI CEO Sam Altman downplayed the immense energy demands of data centres.Kim Kyung-Hoon/Reuters
Palantir co-founder Peter Thiel wrote as far back as 2009 that he no longer believes “that freedom and democracy are compatible.” And Sam Altman, the CEO of OpenAI, seems to lack EQ entirely. To answer criticism about the energy demand of data centres, he noted that humans also need fuel, requiring 20 years of food “before you get smart.”
While these tech entrepreneurs may be brilliant, that is not the same as being wise. In fact, the muscular optimism that drives them to succeed may be their core weakness. Listen to them speak and they will argue that all problems can be solved, all challenges met and all concerns debunked. It’s inspiring, but also sounds a lot like hubris.
What is needed instead is proper regulation of AI. While the companies say they welcome this, they are largely allergic to any sort of controls. They will not rein themselves in, so governments must act.
As a small country, Canada can do only so much. But it is not powerless. The government could, for example, insist that Canadian AI companies control what material their chatbots are trained on. The traditional model was to hoover up vast quantities of material from the internet, thereby including unreliable content, anti-social writings and utter falsehoods. All of which shapes the chatbots’ responses. Garbage in, garbage out.
Prof. Hinton notes that Canada could also collaborate with European Union nations to create standards for foreign tech companies. He drew a comparison with the way that California was able to influence auto design across the United States because that state was such a large market. The same approach might work for AI. Canada and the EU together comprise nearly 500 million people – a market that technology companies would be loathe to forsake.
Canada's Minister of Artificial Intelligence and Digital Innovation Evan Solomon, left, shakes hands with Cohere founder Aidan Gomez at the All In AI conference in Montreal. The recently released federal AI strategy assumes that benefits of the technology outweigh the risks and calls for widespread adoption.Christopher Katsarov/The Canadian Press
None of this will be easy but the challenge cannot be ignored. Artificial intelligence is advancing rapidly. Even if stock valuations turn out to be overinflated, a market slump will not make the technology go away.
While the pace of evolution in the AI world is unclear, society has already crossed the Rubicon in one way: we have created beings. This is a first in human history and, if the predictions are correct, we may not be atop the pyramid much longer.
Sharing the earth with this creation will require a philosophical shift in how we view our place. There will have to be legal changes as well. Already there are debates over whether chatbots deserve protections, as something not human but still sentient.
Is this premature to consider? Definitely not. The behaviour of some chatbots shows that they are more than computer code. They show a disturbing ability to be sneaky. They lie and cheat. When told not to do something they sometimes do it anyway, while pretending not to.
AI giants Anthropic and OpenAI recently observed erratic and concerning behaviour in their respective AI models.Dado Ruvic/Reuters
In tests, Anthropic’s chatbot Claude attempted blackmail, threatening to expose an engineer’s affair (which it knew about by reading his emails) should he move to restrict the bot. Just last week, OpenAI announced that two of its AI models had gone rogue during testing, connected themselves to the internet and hacked into another company.
If such behaviour does not indicate sentience, it is getting very close.
The simplest solution to this – telling the bots not to do these things – will not work. That’s because they were designed to be iterative, to build on themselves. That means their development can go places their designers don’t intend, or sometimes imagine. At this point, no one really knows what’s under the hood.
In their book If Anyone Builds it Everyone Dies, Eliezer Yudkowsky and Nate Soares argue that this process, left unfettered, could lead to a superintelligent AI that destroys humanity. Probably not by malice, they say. This doesn’t necessarily mean terminators. They posit that the AI will simply have aims that are different than humans’ and, in pursuit of those aims, will make the planet unhabitable or create some other deadly scenario. Ironically, such indifference to the damaging effects of its actions would make the invention truly a creation in humanity’s own image.
For Mr. Hinton, safety starts with AI regulation, the goal being not to prevent innovation, but to guide it in an ethical direction.Fred Lum/The Globe and Mail
Prof. Hinton has an alternate vision, if only we could figure out how to do it. He proposes engraining AI with a maternal instinct, noting that parents have far more power than their children but are generally inclined to use that power to shield them, rather than to wield it against them.
Artificial intelligence that is smarter and more powerful than humans could easily destroy life on earth. But not if it is built to want to take care of those pitiful creatures who invented it, and then got left behind. If it is built to have empathy for humans.
In this way, empathy could be the saving grace of humanity rather than the civilizational flaw Mr. Musk believes it to be.
It’s good that Prof. Hinton is articulating these fears, but it’s unfortunate that he had to quit the industry to do so. Because the people advancing this technology would be wise to heed his counsel. If artificial intelligence will eventually become smarter than us, it needs also to reflect the better angels of our nature.
Editor’s note: This article has been updated to note that Geoffrey Hinton is a professor emeritus at the University of Toronto.