Throughout history, specific technologies have revolutionized the way we live, work and play. This has often occurred gradually, but occasionally through dramatic, sweeping changes in society. The technology behind the industrial revolution -- the steam engine, the spinning jenny and the cotton engine, for example -- accelerated economic growth by an order of magnitude, creating an explosion of innovation that still reverberates around the planet.
By contrast, other technologies have had more modest impacts. The invention of electricity, cheap air travel, computing, and so on have all been important and revolutionary but the economic growth they allow has always been constrained by other factors that prevented the kind of explosive growth of the industrial revolution.
That raises an interesting question about the impact of artificial intelligence. This novel technology is currently taking the planet by storm, with no shortage of hype about its potential to change the nature of work. But will this revolution lead to explosive economic growth or will some other factor limit the ultimate impact that AI can have?
Bottleneck Science
Today we get an answer of sorts thanks to the work of Ege Erdil and Tamay Besiroglu, both at Epoch, a research group funded by Open Philanthropy, investigating and forecasting the development of advanced AI. They study the ways AI can produce explosive growth and also the mechanisms that might prevent it.
“We conclude that explosive growth seems plausible with AI capable of broadly substituting for human labor, but high confidence in this claim seems currently unwarranted,” they say, adding that potential bottlenecks need to be better understood.
The main reason why AI could lead to explosive economic growth is that it changes the nature of work. At the moment, economic growth is possible by reinvesting in and improving the factors that lead to growth in the first place. These include land, energy, computing power, access to capital, human labor and so on.
Economic activity requires all these factors and by reinvesting, it is always possible to buy or redevelop land, build more power stations, increase computing performance and so on. Economists call these accumulable resources.
But human labor is non-accumulable because it is limited by demographic factors like the birth rate and immigration. So the availability of human labor places fundamental limits on how quickly economic growth can occur.
But artificial intelligence changes this calculus as soon as it becomes capable of replacing human labor. At that point, it becomes as easy to double or triple a workforce as it is to double or triple computing power or land or energy consumption. You just buy more of it, sit back and watch the explosive growth it allows.
Economists are fascinated by this idea. But many have pointed out that there may be other factors that will prevent explosive growth. So Erdil and Besiroglu have examined these arguments to assess how likely they are to put a spanner in the works.
Many of these arguments, they say, seem highly unlikely to prevent explosive growth. One argument is that there are fundamental limits to growth imposed by the laws of physics and the limited resources of our planet. For example, there are limits to the rate at which information can flow from one volume of space to another and a limit to the amount of energy we can extract from the environment.
In answer, Erdil and Besiroglu say there is no question that these are real and important limits but that we are so far away from them that they cannot represent significant bottlenecks for the foreseeable future.
For example, the Sun currently bathes the Earth in power to the tune of around 1016 W. But we currently use significantly less, roughly 1013 W. So we have three orders of magnitude to play with along with numerous efficiencies we can implement. Access to So power is unlikely to stymie economic growth anytime soon.
Another possibility is that humans might decide they want human-produced goods and so refuse to engage with AI products, thereby limiting growth. In their analysis, Erdil and Besiroglu admit that there may be a preference for certain human-made goods, like works of art. But for the most part, humans will be unlikely to care if computer chips or mobile phones or cars are made entirely by intelligent machines, particularly if they are significantly cheaper.
But other arguments are harder to dismiss. One is that achieving human-equivalent performance is going to be slower than expected and that this will limit the growth it can produce. This argument suggests that some tasks will be easy to automate but that others will be much harder and require orders of magnitude more investment.
An example is that it may turn out to be much harder than expected to build robots with human-like capabilities and even harder to incorporate them into society safely and productively.
But even if this were the case, it is likely that the problem will be solved eventually. So, this kind of problem will delay explosive growth rather than prevent it. “We expect it to be unlikely that this objection blocks explosive growth,” say Erdil and Besiroglu.
One of the biggest potential roadblocks is the problem of aligning AI systems to human expectations and goals. AI alignment is “the challenge of steering artificial intelligence systems to behave according to intended goals and avoid unintended harmful behaviors.”
It’s already clear that current AI systems suffer from problems like hallucinations that need human supervision to correct. Other problems related to safety may emerge with robots and so on. So human oversight might always be required. And that this will create a bottleneck that prevents explosive economic growth.
Erdil and Besiroglu point out that a big factor is the significance of the problems that emerge. AI systems are already becoming better at avoiding hallucinations and so may also be able to overcome other problems. But it’s hard to gauge how these factors will play out in the longer term. “Overall, our assessment is that this argument is most likely not going to block explosive growth, but its influence cannot be ruled out,” they say.
Law Maker Conundrum
Perhaps the most significant factor that could prevent explosive growth is regulation. Policymakers the world over are currently racing to legislate to prevent artificial intelligence being used by malicious actors and to protect consumers. But regulation could also be used to protect jobs and guarantee employment for humans.
Regulation might also limit the material that AI systems can use for training, for example, by preventing AI systems using professionally prepared text or images, without significant compensation to the owners and thereby reducing the incentive to invest in AI systems. Regulation could limit the use of AI in specific activities such as in courtrooms and medical settings or prevent the technology being exported to prevent competition.
The researchers look to history to see how regulation has influenced the spread of previous technologies. They point out that much of the technology for the Industrial Revolution was developed in England in the 18th and 19th centuries.
“While England attempted to forestall the diffusion of some key Industrial Revolution technologies by prohibiting the emigration of skilled workers and the export of machinery, these protectionist policies proved largely ineffective,” say Erdil and Besiroglu. Nevertheless, they conclude that regulation has up to a 1 in 3 chance of preventing explosive economic growth.
Given all these potential problems, the authors say the chances of AI-enabled explosive growth by the end of century is not a foregone conclusion by any means but is about as likely as not. And they say it could start soon if progress in AI continues at its current pace.
That’s interesting work with profound implications. If the analysis proves broadly correct, the importance of steering this economic force toward enhancing human flourishing cannot be overstated, particularly given that the outcome is highly sensitive to the choices we make now and in the years ahead. Policymakers take note.
Ref: Explosive growth from AI automation: A review of the arguments : arxiv.org/abs/2309.11690
