In the words of Jason Furman, the Chairman of the Council of Economic Advisors of the Obama administration: “I want to start with the biggest worry I have about [AI]: that we do not have enough of AI. Our first, second and third reactions to just about any innovation should be to cheer it—and ask how we get more of it.”
This powerful statement from Furman, is a strong signal of the inclination towards a more frequent use of AI by policy makers. By policy making, I refer to the synthesis of complex ideas drawn from within the government community, with the collaboration of experts and the public, in order to design a specific course of action. In practice, economic intuition alone is proven insufficient to capture the complexity of the real world, because no matter how sophisticated the economic model is, it will fail to account for all the phenomena that might be of relevance to design or execute a public policy. Therefore, it is safe to assume that the efficiency of public policies depends fundamentally on how well informed the decision maker is. Nowadays, the main challenge facing most governments is not the scarcity of available data, but rather the inefficient use of it. Although far from optimal, governments have made a considerable effort to overcome this problem. For instance, the United States introduced recently to its public sector the position of Chief Data Officer, whose primary mission is to gather, standardize, manage, and communicate datasets to different public organizations. Now the remaining challenge is how to use this valuable data to tailor the optimal policy, more importantly, how would the use of AI help us achieve such a target?
In principle, policymaking and machine learning share the same doctrine: they both rely on the available information and try to predict a certain outcome. By showing to the machine historical data on individuals’ characteristics and their respective behaviour for the outcome of interest, we train it to predict the future performance of individuals based on their characteristics. The powerful prediction ability of AI is a very precious tool for policy makers, as it allows them to target more efficiently the potential beneficiaries. However, what is truly impressive is the variety of fields in which machine learning could contribute. For example, machine learning could help policy makers predict more accurately which students are most likely to dropout of schools based on their socio-economic characteristics or on their previous educational performance. Also, it could be used to predict which patients are most likely to have heart attacks, or to be diagnosed by cancer in the future, based on their medical history and their genetic codes.
If artificial intelligence is so proficient in assessing information and predicting future outcomes, at least better than humans, shouldn’t we just replace economists with machines? At least we might avoid another humiliating Michael Fish moment. As a matter of fact, economics as a profession is more and more criticized for its low ability to forecast and prevent crisis, especially after the financial crash of 2008. Thus, some would argue that AI is a better substitute to human economists, especially when replacing humans by machines has proved to be very efficient in other professions such as weather forecasting.
Whilst it is true that relying on AI might be optimal for economic forecasting, trusting it exclusively for policy making is extremely risky. For instance, the performance of AI is as good as the data it is modelled on. If the data is biased in any way, all the policy implications that follow will be also biased. In other words, there exists an irreplaceable human factor that takes into account all relevant social, political, moral and ideological issues at stake, which AI and machine learning simply fail to do. In short, it is immature to think of AI as a substitute for policy making. However, it is a valuable tool that must be used more frequently by policy makers as it gives them much better predictive abilities, unreachable otherwise. Finally, it allows policy makers to have a better understanding of the population that they serve, which is crucial to design and implement the optimal policy.
By Omar Doghiem