Artificial Intelligence Still Isn’t All That Smart
Artificial Intelligence Still Isn’t All That Smart
In the business world, machine learning often goes by the annoying moniker of “artificial intelligence.” That science-fiction buzzword evokes visions of godlike sentient computers, when in fact, the product is much closer to a statistical regression. Machine learning is about using algorithms to predict things — whether a web-security image contains a cat, what a Google user wants to search for, or whether a self-driving car should brake to avoid a crash. No one yet knows how to give a single computer system the mental flexibility to reason and learn like a human being.
But buzzwords or no, the field is hot. AI startups have been getting more and more funding in recent years.
Big tech companies such as Alphabet Inc. (Google), Apple Inc. and Amazon.com Inc. are investing heavily in the technology. Starting salaries for specialists in the field can be as high as a half-million dollars. AI startups are being acquired at a prodigious rate. Corporate incumbents are often wary of creating the technologies that are sure to disrupt their existing business models, but these giants are throwing caution to the wind.
It’s important to note that machine learning hasn’t yet made its mark on the economy — to paraphrase economist Robert Solow, you can see the machine learning age everywhere but in the economic statistics.
Employment levels have returned to healthy levels, and there’s no evidence that machines are taking many of our jobs yet.
And productivity is only rising at a slow, if steady, pace.
Nor is the AI investment boom yet big enough to cause a general investment boom.
But like the computer revolution three decades ago, the machine-learning revolution will eventually make an impact. The question is how.
A few economists have tried to provide some preliminary answers. In their book “Prediction Machines: The Simple Economics of Artificial Intelligence,” Ajay Agrawal, Joshua Gans and Avi Goldfarb — three researchers who have made the economics of machine learning their specialty — try to provide an answer. Since the technology is both so enormously broad and so new, their answer is inevitably highly speculative and general — consider trying to predict the implications of the steam engine in 1800, or the internet in 1992. But considering the scope of their task, Agrawal et al. do an excellent job.
Agrawal et al. deal with machine learning in the classic econ way — by first defining what the product is, then thinking about the supply and demand for the product. As they see it, the technology represents an increase in the supply of predictive power — any task that relies on making predictions (very generally defined) is now going to be easier and cheaper. But these economists aren’t just spouting theory. Having worked with the Creative Destruction Lab, a startup incubator at the University of Toronto’s Rotman School of Management founded by Agrawal himself, the authors are able to draw on a wealth of real cutting-edge examples to bolster and illustrate their general claims.
The authors generally don’t envision a world of full automation, with machines replacing humans at every step of the production process. Instead, they see machine learning being deployed selectively at some nodes of the value chain where data is plentiful, leaving human judgment to focus on the rest. Though “judgment” is a fuzzy word, Agrawal et al. basically identify two cognitive tasks in which humans will beat intelligent algorithms for the foreseeable future — making predictions based on small data samples, and identifying what constitutes success and failure. Humans are still better at knowing what they want, and at modeling the underlying structure of how the world works.
If the authors are right, machine learning will revolutionize white-collar jobs in much the same way that engines, electricity and machine tools revolutionized blue-collar jobs. Just like machine tools allow workers to skip some physical tasks and apply their muscle power only to a few essential things, machine learning tools will allow workers to skip some mental tasks and apply their brain power to only a select few things. The result, presumably, will be another big increase in productivity.
On the downside, if the set of tasks amenable to machine learning increases at a very rapid clip, it could leave lots of workers displaced and adrift as they are constantly forced to change their job description. Machine learning could also accelerate the trend toward the elimination of routine jobs, hollowing out the middle class and increasing inequality.
This is a very simplified picture of the economics of machine learning — in reality, economics is about much more than supply and demand. In a new paper, legendary economist Hal Varian — who helped Google design the pricing system that underlies its remarkable profitability — briefly addresses some of the thornier questions of how smart machines will change the industrial structure of advanced economies. Not all of these are salutary. Machine learning-enabled price discrimination might allow companies to figure out exactly how much customers are willing to pay for things, and gouge them for every penny. Machine learning-enabled network effects could aggravate the problem of big-company dominance. Algorithms could collude with each other to rig markets, without humans even realizing the collusion is happening.
In the end, the only way to really know what machine learning will do to the economy is to wait and find out. It’s important to anticipate and be ready for possible problems that might arise, but just as was the case with computers, the internet and industrial technology itself, machine learning’s ultimate effects will probably surprise all of us.