Artificial intelligence and public policies

Most economists believe that artificial intelligence (AI) will have a major effect on the economy.

Mr. Yassine Lefouili, Senior Lecturer at the Toulouse School of Economics (School chaired by the Nobel Prize for Economics Jean Tirole) and Director of the TSE Digital center

This impact will depend to a large extent on several public policies that could be divided into two categories: on the one hand, those affecting the development and dissemination of AI such as investment support policies or the personal data protection policies and, on the other hand, those that deal with the consequences of using AI such as employment policies or competition policies.

AI has two characteristics that require the adoption of specific public policy measures. The first one is that AI affects virtually all the industries, this is why it is qualified of “general purpose technology” just like the steam engine, electricity, or microprocessor. This peculiarity must be taken into account when developing policies to support investment in AI as the private sector tends to adopt a mechanism that leads to underinvestment in the technology.

To illustrate this mechanism, let’s consider a start-up that is active in the field of facial recognition. Imagine that, through its investments in research and development (R&D), the start-up in question makes improvements to machine learning algorithms that are also used in the field of medical diagnosis. These improvements certainly benefit the company but they also benefit companies active in the field of medical diagnosis, directly if they can use the same enhanced algorithms or indirectly if they can draw inspiration from the improvements made to come up with new ones. Here we have an illustration of what economists call a “positive externality”, that is, a situation where an action taken by one economic agent benefits other economic agents.

The above example illustrates a more general phenomenon, namely the existence of positive cross-industry externalities resulting from the fact that investments in a given AI application can improve the underlying AI technology and thus benefit consequently makers and users of other applications of this technology. These externalities are generally not integrated into the economic calculation of a company which, when faced with an investment decision, does not take into account the total value it creates but only the part of this value that it is able to capture. Consequently, companies will tend to invest less than what would be desirable from a collective point of view. This phenomenon of underinvestment is not unique to general purpose technologies such as AI, but is accentuated in this case by the existence of significant cross-sectoral externalities. This justifies the adoption of more ambitious investment support policies than those meant for technologies less concerned by such externalities.

The second characteristic of AI that should be emphasized is the fact that it could radically change the very nature of the R&D process. This is why some economists consider AI a new research tool and some go as far as to call it a new method of invention.

The first implication of using AI as a research tool is that there will certainly be a change in the division between capital and labor in the field of R&D, with a decline in the share of labor that will profit the share of capital. In particular, it is likely that tasks of a repetitive nature currently performed by researchers or engineers, such as searching and classifying scientific information, will increasingly be assigned to algorithms. As highly qualified labor is a scarce resource and therefore a barrier to entry into the R&D market, it is possible that the use of AI, by partially destroying this barrier, could lead to an increase in the number of companies active in the field of R&D.

This optimistic vision of an AI leading to a form of democratization of R&D is unfortunately incomplete. Indeed, the use of AI as a research tool can also be accompanied by the establishment of a new barrier to entry if access to the training data necessary for the functioning of the algorithms is limited. Such a scenario can occur for two reasons. First, a dominant player may be tempted to restrict access to its training data in order to protect its leadership position. Second, it is possible that companies would like to engage in data transfers but cannot do so effectively due to the absence of an organized market in which such transactions could take place and the associated legal risks that could arise.

Public intervention is desirable in both of these situations. In the first case, public authorities should not hesitate to use the tools of antitrust policies to put an end to possible anti-competitive practices of the dominant actor or, even better, to dissuade them. In the second, they must encourage the emergence of an environment conducive to the transfer of training data, in particular by providing a favorable legal framework for such transactions.

(*): Mr. Yassine Lefouili, Senior Lecturer at the Toulouse School of Economics (School chaired by the Nobel Prize for Economics Jean Tirole) and Director of the Digital Center (TSE). His main areas of research focus on competition policies and the economics of intellectual property. His research has been published in several reference journals such as the RAND Journal of Economics.