Note: Thanks to Tim Blomfield and Nicholas Emery for contributions to the literature summaries and organization.
Introduction
At least since Herbert Simon’s 1960 prediction that artificial intelligence would soon replace all human labor, many economists have understood that there is a possibility that sooner or later artificial intelligence (AI) will dramatically transform the global economy. AI could have a transformative impact on a wide variety of domains; indeed, it could transform market structure, the value of education, the geopolitical balance of power, and practically anything else.
We will focus on three of the clearest and best-studied classes of potential transformations in economics: the potential impacts on output growth, on wage growth, and on the labor share, i.e. the share of output paid as wages. On all counts we will focus on long-run impacts rather than transition dynamics. Instead of attempting to predict the future, our focus will be on surveying the vast range of possibilities identified in the economics literature.
The potential impact of AI on the growth rate of output could take the form of
- a decrease to the growth rate, even perhaps rendering it negative;
- a permanent increase to the growth rate, as the Industrial Revolution increased the global growth rate from near zero to something over two percent per year;
- a continuous acceleration in growth, with the growth rate growing unboundedly as time tends to infinity; or even
- an acceleration in the growth rate rapid enough to produce infinite output in finite time.
Basic physics suggests that the last of these scenarios is impossible, of course—as is eternal exponential or super-exponential growth. The relevant possibility is that AI induces a growth path that resembles these benchmarks for some time, until production confronts limiting factors, such as land or energy, that have never or have not recently been important constraints on growth. We will thus also consider how increases in growth may eventually be choked off by such limiting factors.
The potential impact of AI on the labor market includes predictions that
- wages fall;
- wages rise, but less quickly than output, producing a declining labor share;
- wages rise in line with output, producing a constant labor share; or even
- wages rise more quickly than output, producing a rising labor share.
In the respective limiting cases, AI could result in a future in which wages are (literally or asymptotically) zero or near zero; very high in absolute terms but approximately zero percent of total output; or high both in absolute and in relative terms.
As this discussion illustrates, the space of possibilities is vast. At the same time, as Simon’s failed prediction testifies, transformational impacts from AI are by no means certain to transpire on any particular time horizon. Most studies to date of the economics of AI, therefore, have focused on the most immediate, moderate, and likely impacts of AI. These include marginal shifts in output and factor shares; impacts on marketing and statistical discrimination; impacts on regional inequality; and industry-specific forecasts of AI-induced growth and labor displacement over the next few decades.
Empirical estimates of the future economic importance of AI have drawn inferences from foreseeable industry-specific applications of AI, from totals spent on AI R&D, and from comparisons between AI and past technological developments in computing, internet connectivity, or related fields (see e.g. Chen et al. (2016)). Industrial and R&D-based inferences may however severely underestimate the field’s transformativeness if technological development has substantial external effects, as is often assumed; the importance of the atomic bomb is not well approximated by the cost of the Manhattan project. Even in the absence of externalities, furthermore, inferences from R&D expenditures implicitly discount future output according to the time preference of the research funders, who may assign negligible value to the impact of their technologies on the distant future. And common reference class-based projections preclude the possibility that AI ultimately proves to be truly transformational, less like the economic impact of broadband than like that of the Industrial Revolution or the evolution of the human species itself.
In recent years, economists have begun to engage earnestly in formal theoretical explorations of a wide array of the transformative possibilities of AI, including those outlined above. We aim to summarize the findings of these explorations. In the process, we hope not only to state the conclusions of various models, but also to give the reader some mathematical intuition for the most important mechanisms at play. (Indeed we have altered some of the models slightly, to clarify their implications regarding transformativeness and their relationships to the other models discussed here.)
This document is intended for anyone comfortable with a moderate amount of mathematical notation and interested in understanding the channels through which AI could have a transformative impact on wages and growth. Readers with backgrounds in economics will hopefully come to better understand the possibilities which concern singularitarians, and readers with singularitarian backgrounds will hopefully come to better understand the relevant tools and insights of economics.
The rest of this document proceeds as follows. §2 consists of an overview of the economics that will be relevant for understanding the subsequent sections. §3 discusses models in which AI is added to a standard production function. §4 discusses models in which AI is added to a “task-based” production function. §3 and §4 both implicitly take place in a setting of exogenous productivity growth; §5 discusses models in which productivity growth is endogenous and AI can feature in its production. §6 compares the results found in §3-5. Finally, §7 concludes.
