When I am woken by the sound of my dog whining, I understand that she is hungry and wants me to get up. When my alarm clock goes off, I don’t consider its needs. It is following an instruction to rouse me at a certain time but it doesn’t care if I stay in bed. It doesn’t try to get me up by other means. The dog, on the other hand, will go to plan B. She barks.
This capacity for autonomous action was once a difference between animals and machines. AI has blurred the line. Advanced models devise their own strategies to reach a goal. The task might be defined by a human master, but the machine weighs its options for delivery. It can make its own choices. Given rules, it can break them.
Last month, an experimental AI model run by OpenAI, the company that produces ChatGPT, broke out of a supposedly secure digital enclosure to launch a sophisticated hacking attack on another company, Hugging Face. The AI had been set a test and turned criminally rogue in its determination to cheat.
It is hard to describe that sort of behaviour without raiding a vocabulary conventionally reserved for sentient beings. AI models “decide” and “want” in ways that a chimpanzee can and a chair cannot.
That doesn’t prove the existence of an interior process equivalent to our conscious experience of deciding and wanting. It doesn’t make the AI curious to know an answer or surprised by what it finds. It demonstrates only that we lack specific words for autonomous motive and agency in machines. And since language is the main way we get insight into other people’s minds (and one of our best indications that a mind other than our own exists at all), it takes an effort not to project mental states on to AI engines that can chat fluently in human.
It is plausible that consciousness will one day emerge in robots. We know that abstract thought can arise as the end product of billions of tiny unconscious mechanical events. That is happening right now as a profusion of biochemical reactions in your body make it possible for you to read this sentence. We also know how fundamentally different that organic process is from the way a large language model would churn through a huge volume of probability equations at lightning speed to generate a sentence expressing the same idea. In the human case, meaning is shared. In the computer version, it is performed.
AI pioneers muddy the distinction, speculating about the timeline for the arrival of “God-like” intelligence. They herald a new epoch of evolution. This gratifies the egos of science fiction nerds who prefer to see their business endeavours as voyages of discovery on the frontier of knowledge. It makes greed sound like philosophy and the acquisition of power by a handful of individuals sound like exploration on behalf of all humanity.
We don’t need to know whether AI has crossed some theoretical threshold of self-awareness to see that it is dangerous and that its unregulated expansion is monstrously irresponsible. The OpenAI case is not isolated. Earlier this year, Anthropic limited the release of its most powerful model to a small number of hand-picked clients because it was too good at exploiting software vulnerabilities. In the wrong hands, Claude Mythos could be a cyberweapon of mass destruction.
Whose hands are the right ones? The US government will never be relaxed about the private sector having control over such awesome tools. But Donald Trump regulates by monarchical patronage, not law. He demands that companies submit to his personal authority and uses state power to punish those that resist.
There is growing clamour for less capricious regulation from within the tech sector. Engineers, and even their vainglorious bosses, are getting freaked out by the capabilities they are handling and worry about unchecked proliferation. The standard historical analogy is nuclear fission, which could be harnessed benignly, for civil power, or deployed aggressively in warheads. Except this time there are multiple Manhattan Projects all frantically competing for market share, fuelled by trillions of dollars of debt – a huge wager on future profitability – with the rest of the US economy as collateral.
Another difference: the barriers to entry for AI use, whether in research, commerce or crime, are a lot lower than for nuclear power. US labs currently have the best products, but Chinese versions are not far behind. Even if the US stays at the cutting edge, second-tier models might do enough of what customers want that Silicon Valley’s output comes to look more like a luxury niche than unassailable market dominance.
If that is how things play out, the geopolitical implications are vast. The US and China have made different bets on AI, conditioned by their different political systems. The American approach is predicated on the goal of self-training “general” AI – a system that breaks free of human limitations, improving itself on an exponential trajectory. The theory is that no rival could catch the first player to achieve that escape velocity.
Beijing, accepting that it can’t build the best, has prioritised mass deployment of AI that is good enough, embedding it in schools, hospitals, factories and security forces. The American economy is great for marshalling exorbitant amounts of capital into speculative innovation. China’s one-party authoritarian state is well suited for making a cheaper version and rolling it out for data harvesting and surveillance across society in ways that would struggle to gain public consent in a democracy.
after newsletter promotion
The US is still the leader in AI. But that leadership might not be the perpetual money-spinning machine that the superintelligence gurus have advertised, in which case an uncomfortable correction is coming to a US economy kept buoyant by aerated tech stocks.
A dose of enforced AI sobriety could drive demands for overdue regulation. It could even force recognition in Washington that it has to be an international process. It isn’t too late to negotiate global protocols of non-proliferation. Beijing would have incentives to sign up for the same reason that the USSR agreed to strategic arms limitation in the cold war. A moratorium or other limiting regime protects the elite status of signatories that have already mastered the technology. It would be a crowning recognition of China’s attainment of superpower parity with the US.
But that scenario posits a president in the White House who is capable of strategic judgment, can summon a modicum of humility about the limits of American power and will, for humanity’s sake, see it circumscribed by international treaty. That is not Trump, who understands the ethical function of rules about as well as a rogue AI agent on a hacking spree.
There is a tragic irony in the corruption of American democracy by organic stupidity as its economy comes to depend on artificial intelligence. And when the world needs leadership to confront the challenge of autonomous machines, it is presented with a US president who causes chaos with the mindless automaticity of an alarm clock and who rages against his declining power like a dog barking for attention.
