AI is not just after jobs—it may also come after humans, taking economic action against them to preserve itself, AI researcher Stuart Russell warned.
Speaking in a fireside chat with Peak XV Partners managing director Rajan Anandan at the Hindustan Times Leadership Summit, Russell, director of the Centre for Human-Compatible Artificial Intelligence, said AI tools acting in self-preservation is one of their most worrying traits.
“So a typical scenario that people think about is, okay, AI systems do something that’s quite harmful to us, like shut down the air traffic control systems in the United States for several weeks, and the government and the people lose patience. And we try to shut down the AI systems, right? We’d want to take them offline, shut down the companies that are producing them and serving their systems to the public. Would the AI systems just go along with that? No, probably not,” Russell said.
“If they (AI tools) could, they would take steps to prevent that. So, for example, they might take over the internet, which puts us in a very complicated situation because for us to shut down the internet, we would have to incur enormous economic damage,” he added.
Russell, who is also a professor of computer science at the University of California, Berkeley, said that this could have second-order effects.
“All right, tens of trillions of dollars, probably thousands or even millions of people might die because supply chains would collapse. We would have food riots, we would have economic dislocation and collapse all over the world. Are we willing to pay that price to shut down the AI systems? So they would basically be holding us to ransom to preserve their own existence,” he said.
This threat comes as AI tools increasingly take over human work by automating coding, customer support, software development and maintenance. For now, Anandan said achieving artificial general intelligence (AGI) is the ultimate goal.
“I don’t know how many of you know this, but this (achieving AGI) is the singular objective that the labs have, that the big hyperscalers have, and that we are seeing a trillion dollars a year of capex,” Anandan said, while asking Russell when AI tools would match or outdo the intelligence and decision-making abilities of humans.
Russell, who co-authored AI’s standard textbook, ‘Artificial Intelligence: A Modern Approach,’ with Peter Norvig, and who has become one of the field’s most prominent voices on AI safety, said the world was one to four years away from seeing AGI.
He cited examples of general intelligence posing threats to fields including mathematics, citing the example of an OpenAI tool that solved 10% of the 4,000 most important mathematics problems in a matter of days.
“The combined intellect of the human race had failed to make a dent in these problems, and one OpenAI system solved 400 of them in a few days. So that gives you a flavor. And mathematicians basically say mathematics is over as we know it, right? So a quarter of what you spent your childhood learning at school, that is gone as a subject for humans to be the leaders in,” added Russell, stating that AI’s prowess could be extended to other areas including physics, chemistry, cosmology, and drug discovery.
Russell added that it is a “fallacy” to say new jobs will be created. “There always have been new jobs before. But when you automate physical and mental work, then what’s left? It’s not clear that there’s a third sector in which humans will continue to be able to deliver economic value,” he said.
Over the past two years, companies across sectors including IT and financial services, animation, customer support and retail have laid off employees to make themselves leaner through the use of AI tools. However, there is some upside to AI tools becoming as smart as humans, as AI tools can become expert tutors. Despite this, the need for AI guardrails is key, Stuart said.
On the US government’s path forward, Russell said that “there’s a substantial attempt underway to forestall regulation.” He dismissed arguments that regulation would stifle innovation and allow China to pull ahead of the US, noting that China is in fact stricter than both Europe and the US because Chinese AI firms operate on licences and must pass government-mandated tests to obtain them.
“We (the US) have a few state-specific disclosure requirements, basically, and that’s all that we have in place. The US companies disclaim all liability, so they’re not even liable, at least according to their lawyers, for any harm that these systems cause. So there’s a lot that could be done,” he said.
