You can barely listen to a podcast or read the news at the moment without hearing an apocalyptic warning about artificial intelligence. The head of Microsoft AI suggests the technology could become a new 鈥渟ilicon species鈥� that . A former Anthropic employee says there is a at the hands of it. And Geoffrey Hinton, the 鈥済odfather of AI鈥�, says we may lose control of it entirely . Is AI really out of control?
, formerly of the Alan Turing Institute and now co-founder at non-profit company Our AI Collective, warns against accepting predictions based on the reputation of those making them. Instead, we should look to facts 鈥� and the facts don鈥檛 suggest that AI has broken free from human control or that it is on the cusp of doing so.
鈥淭hat鈥檚 a narrative that gets pushed again and again. We get these waves of hysteria,鈥� says Aitken. If anything has gone rogue, she says, it isn鈥檛 AI models, it鈥檚 the humans developing them.
When AI models go and hack another company鈥檚 computers, it is because appropriate safety measures weren鈥檛 put in place during testing 鈥� testing that was designed to assess how well those AI models could hack into computers. We are simply seeing the failings of the companies to erect adequate safeguards.
Advertisement
Nor, says Aitken, are AI models improving themselves recursively 鈥� indicating that we could be approaching the dreaded singularity where AI slides into a self-driven loop of improvement that accelerates and quickly creates artificial general intelligence. 鈥淲hen these ideas of superintelligence are put forward, they鈥檙e based on a compelling sensational narrative, but there鈥檚 no concrete evidence,鈥� says Aitken. 鈥淚f there was something as powerful as that, we would have seen something of it.鈥�
at the University of Surrey, UK, says that, in truth, there has been no dramatic shift in the risk posed by AI over the past couple of years, despite its advances in mathematics and cybersecurity. But hype is periodically stirred up by an incident or a catchy soundbite.
鈥淚t鈥檚 very easy to create this sort of frightening story. It sort of plays to the human psyche,鈥� says Rogoyski. He believes that in large language models, humans have created something that is apparently so similar to themselves that it holds up a mirror. 鈥淭hen we imbue it with all of the human-like behaviours and fears. So we鈥檙e terrified of it,鈥� he says.
If all that is true, then why are so many tech bosses calling for a slowdown in AI research?
Aitken sees this talk as a way to head off law-makers from creating regulations that would hinder AI firms鈥� work 鈥� by seeming to act responsibly, they lessen the demand for government to step in to force the issue. But Aitken points out that when former staff make bold predictions about superintelligence and the risk posed by AI models, these same companies do little to shut things down. 鈥淭hese superintelligence narratives are actually helpful to the AI companies because they build hype, they build excitement, they boost investment,鈥� she says.
The result is that the companies end up walking a narrative tightrope: reassuring us all that they are acting ethically and responsibly by calling for a slowdown that is unlikely to happen, yet playing up to sci-fi ideas of superintelligence because it is good for business.
Rogoyski believes another factor could be further muddying the waters. Gargantuan AI investment in data centres, staff, research budgets and training data is simply unsustainable, so an official slowdown could give the companies a way to slow the pace without losing out to competitors 鈥� perhaps even cementing their advantage.
It could be that companies know that further scale isn鈥檛 important for day-to-day business tasks like sending emails, filing reports or designing social media graphics (although it may be key to unprofitable endeavours like solving mathematical problems), so they would actually relish the chance to ease off and begin to make efficiencies, cut costs and increase revenue.
鈥淒o I really need to spend a trillion dollars in order to be able to write a better PowerPoint slide or get some marketing emails out? It really doesn鈥檛 stack up, and we鈥檒l look back on this in 10 years and say, 鈥榃hat were we thinking?'鈥� says Rogoyski.
But that leads to another risk posed by AI, which is far more concrete and less theoretical: a bursting of the AI bubble, and possible economic fallout. Rogoyski expects less of a dramatic 鈥減op鈥� and more of a gradual 鈥減fffft鈥� as the gas leaks out of the industry in coming years.
Among all this, however, there are legitimate concerns for the future. Rogoyski says there is a real 鈥� but small 鈥� risk from AI, which needs to be taken seriously. Whether the model that becomes advanced enough to pose that risk arrives soon, or many years and some fundamental breakthroughs from now, it is worth planning ahead.
Rogoyski says he doubts that any current AI is intelligent, although our definition of that word and understanding of its mechanisms are poor. But when it comes, we shouldn鈥檛 expect it to look or act like us: 鈥渢here will be patterns of thought and behavior that will be very alien鈥�.
So to try to spot that kind of intelligence is even harder, and to try to predict what its motivations might be, harder still.
鈥淚f you鈥檙e building a machine that you believe is going to be intelligent, then you ought to be having a discussion as to what rights it should have,鈥� says Rogoyski. 鈥淒oes it mean I could just have an off switch that, when I get bored a bit or I feel threatened by it, I can kill it? That doesn鈥檛 seem to be terribly fair, nor necessarily helpful, because machines that feel threatened by humanity might behave in a different way. It might become a problem in itself, to not grant it rights.鈥�
