In my last two columns, I argued that AI brings both opportunities and great dangers, some even existential. This transformative technology also threatens fundamental values, including personal and institutional accountability, the rule of law, democracy and even what it means to be human. Moreover, AI will be hard to regulate successfully, not only because its impact will be pervasive, but because progress is being driven by fierce competition among businesses and between the US and China.
Remarkably, a recent post from Anthropic states that “we are delegating a growing share of AI development to AI systems themselves . . . Taken far enough, and given enough compute, that trend points to an AI system capable of . . . autonomously designing and developing its own successor.” The post then states that “if it were possible to . . . slow the development of this technology, to give ourselves more time to deal with its immense implications, we think that would likely be a good thing”. If even Anthropic, a leader in AI, is fearful of what lurks ahead, the fears of the rest of us, especially the young, can only be reinforced.
A big part of this politically salient concern is over the unemployment downside of the assumed productivity upside. But the speed and scale of the transformation AI will produce is unknown. My colleague John Burn-Murdoch noted recently, for example, that the increase in the supply of apps generated by AI has not led to a corresponding rise in their use. It has also generated a bigger jump in output in the early stages of software development than in final products. (See charts.)
Again, an overview of the employment impact of AI issued by the International Labour Organization last year concluded that, globally, one in four workers is in an occupation with some exposure to generative AI. But, it also adds, just “3.3 per cent of global employment falls into the highest exposure category”. This does not seem a huge disruption. Moreover, in the past long lags have occurred between big innovations (electricity, for example) and higher productivity. As Paul Krugman writes, productivity growth has been lower during the digital era than it was after the second world war, a period without such breakthroughs.
At the opposite end of the debate, Vinod Khosla, an experienced technology investor, states in the FT that “I’m certain AI will do 80 per cent of the economically valuable work humans do today, for 80 per cent of all jobs, faster than most believe. The question isn’t whether mass underemployment arrives by the next decade, but whether we have a coherent policy framework ready when it does.”
Scepticism about the speed and scale of AI’s impact is justified. But Khosla is right: we need to prepare. Civilisation might not survive the existential shocks and economic disruptions that AI threatens. Uncertainty justifies readiness, not complacency.
So, what must readiness mean?
First, we need to be prepared for a world in which machines will be taking big and in some cases — notably those of war and biological research — hugely consequential decisions. Ultimately, humans have to be accountable for those decisions, as programmers of AI, managers of the companies that sell it and decision makers in institutions that use it. Contrary to Argentine President Javier Milei’s view, AI must not run institutions without accountability of people. Owners, managers and officials should be liable to criminal and civil penalties for damage caused by AI.
Second, we cannot rely on the moral sense and self-restraint of AI creators. We have already had terrible experience with social media. As I have noted: “Disseminating lies and frauds can be good business. Worse, disseminating posts that make people’s lives unbearable can be a good business . . . Artificial intelligence seems likely to worsen our collective plight by creating ‘perfect’ frauds of all kinds.” Anthropic may want to slow the pace. But it is in a race: it cannot control what its competitors do. We do not allow drugs companies to release medicines that have not gone through a suitable testing regime, for very good reason. Something similar should apply to new AI software. Moreover, in a competitive business, such regimes also have to apply globally.
Third, this is why regimes cannot just be national. There must be a global agreement on how AI is to be tested and controlled and how liability for damage is to be imposed. The EU, it appears, is once again playing the role of regulator of first (or last) resort. This may not be such a bad thing. People around the world even trust the EU to be a better regulator than the US or China, probably because they believe it will be less captured by business interests or the desire to use AI as a weapon. But ideally, China and the US should be the keystones of any agreement. AI is too risky for everybody to develop unchecked.
Last and particularly important, there is a good chance that AI will in time devastate the labour market, increase inequality and create an extraordinary concentration of economic — and so political — power in the hands of a tiny number of businesses and people. If we add this to the many other threats the technology poses, we confront an enormous risk of an autocratic overthrow of democracy. Indeed, it is already happening. Those who wish to see the survival of government of the people, for the people, by the people must try to stop this. The most obvious implication is that a good part of the increased income and wealth must be shared. The time to prepare for this is now. If we do not act, it will be far too late.