The Economics of Artificial Intelligence
Page 80
elasticity of substitution (CES) production function so that there is love
of variety. At any date t there is a measure N( t) of varieties. The marginal returns to new varieties are positive, but diminishing. The key “building
on the shoulders of giants” externality is that the cost of developing a new
variety is inversely proportional to the measure of varieties. As a result, inno-
vation costs fall over time, generating endogenous growth. A one- sector,
two- country extension appears in Rivera- Batiz and Romer (1991). A two-
sector, two- country extension appears in Grossman and Helpman (1991).
This brief review leads to a number of observations. As in the previous
section, the benefi t of trade policy depends on whether the externality
operates at the national or international levels; Q of the previous section is replaced here by either n( t) or N( t). Hence, if each fi rm builds on the international frontier n( t) or the international number of varieties N( t), then there are no implications for comparative advantage; however, if each fi rm builds
on its national n( t) or national N( t) then the frontier country will develop an increasingly strong comparative advantage in the quality or expanding-varieties sector. With national- level externalities one country will capture
the lion’s share of the quality/ varieties sector. Further, a country can capture
this sector by using R&D and trade policies.
Endogenous growth models provide important insights into the details
of R&D and trade policies. Research and development policies directly tar-
get the knowledge externality and so are preferred to (second- best) trade
policies. One R&D policy avenue is to promote knowledge diff usion. This
can be done through subsidies to nonprofi t organizations targeting local
within- industry interactions and industry- university collaborations. A sec-
ond R&D policy avenue is to promote knowledge creation through R&D
subsidies that are available to all fi rms, universities, and students. There
is a tension between these two avenues; knowledge diff usion can discour-
age knowledge creation since knowledge diff usion to competitors reduces
the returns to innovation. However, the tension is sometimes constructive:
Silicon Valley emerged from the shadows of Massachusetts’ Route 128
partly because of an “open- source attitude” (Saxenian 1994) and Califor-
16. Placing endogenous growth into a two- sector model so as to facilitate a discussion of comparative advantage is not easy because the sector with improving quality slowly takes over the entire economy unless other price or nonprice “congestion” forces prevent this.
Artifi cial Intelligence and International Trade 479
nian restrictions on noncompete clauses (Marx and Fleming 2012). It is less
likely that diff usion of knowledge to foreign countries will be as benefi cial
domestically.
This class of models discourages policies that target individual fi rms
or that “pick winners.” To understand why industry leaders should not be
advantaged by policy, note that counterintuitively, industry leaders will be
the least innovative fi rms due to the “market- stealing” eff ect. If an entrant
innovates, it steals the market from the leader. If a leader innovates, it canni-
balizes itself. Leaders therefore have less of an incentive to innovate. Aghion et al. (2001, 2005) address this counterintuitive result by developing a model
in which leaders innovate in order to escape the competition. Aghion et al.
(2017) and Lim, Trefl er, and Yu (2017) are currently developing international
trade models featuring escape the competition.
In the context of AI, none of the above endogenous growth models is
ideal, leading us to conjecture about what an appropriate model might look
like. The advantage of endogenous growth models is that they emphasize
knowledge creation and diff usion. Thinking more deeply about AI develop-
ment and commercialization, it is useful to distinguish two aspects of what
is done in the AI research departments of large fi rms. First, they improve
AI algorithms, which have the fl avor of quality ladders. (Recall that qual-
ity can be something that is perceived by consumers or, as is relevant here,
something that reduces marginal costs.) Second, AI research departments
develop new applications of existing AI; for example, Google uses AI for
its search engine, autonomous vehicles, YouTube recommendations, adver-
tising network, energy use in data centers, and so forth. This suggests an
expanding- varieties model, but one that operates within the fi rm. We are
unaware of any endogenous growth models that have both these features.
Grossman and Helpman (1991) have the fi rst and Klette and Kortum (2004)
have the second. Combining them in one model is not trivial and analytic
results would likely have to be replaced with calibration.
19.3.4 New Economic Geography and Agglomeration
The discussion in the previous section points to the possibility that
knowledge spillovers are subnational, and this leads naturally to a theory
of regional clusters such as Silicon Valley. New economic geography or
NEG (Krugman 1980) does not typically consider knowledge spillovers, but
it does consider other local externalities that drive regional clusters. Three
mechanisms have been particularly prominent: (a) demand- side “home-
market eff ects,” (b) upstream- downstream linkages, and (c) labor- market
pooling. All of these theories feature two key elements: costs of trading
across regions (e.g., tariff s) and increasing returns to scale at the fi rm level
(which can be thought of as the fi xed costs of developing a new product). We
explain the role of these two elements in the context of home- market eff ects.
480 Avi Goldfarb and Daniel Trefl er
Consider a model with CES monopolistic competition and two regions
( j = 1, 2). There are varieties of machines and the larger the set of machines to choose from, the more productive are the producers. Let N be the measure
j
of machine varieties available in region j. Then with CES production func-
tions, productivity is proportional to N .17 The fundamental factor push-
j
ing for agglomeration is the strength of this love- of-variety/ productivity
externality. (This is related to the externality in Romer’s expanding varieties
model, which is also proportional to N .) As in previous models, the exter-
j
nality operates at the local level rather than at the international level. This
externality encourages fi rms to colocate or agglomerate since the agglomera-
tion of fi rms drives up N and productivity. The fundamental factor pushing
j
against this agglomeration is trade costs: a fi rm can avoid trade costs by
locating close to consumers rather than close to other producers. The main
insight of this model is that in equilibrium a disproportionate share of the
world’s fi rms will locate in a single region, and this region will thus have
higher productivity. As a result, this region will be richer. Notice that fi rms
are choosing to set up where the competition is greatest and where wages
and property values are the highest.
The above model of agglomeration has been extended in countless ways
(e.g., Krugman and Venables 1995; Fajgelba
um, Grossman, and Helpman
2011; Duranton and Puga 2001) and it is easy to think of applications where
the force for agglomeration is not the variety of machines, but the variety
of knowledge held by fi rms. If this knowledge is tacit (meaning it cannot
be codifi ed and transmitted in a document), then knowledge spillovers are
only transmitted locally via face- to-face interactions. In this case, knowl-
edge externalities lead fi rms to agglomerate. The result is regions like Silicon
Valley.
19.3.5 Cluster
Policies
Cluster policies have long been the politician’s best friend, yet economists
remain highly critical of them. In surveying the evidence for the success of
these policies, Uyarra and Ramlogan (2012) write “There is no clear and
unambiguous evidence that over the long term clusters are able to gener-
ate strong and sustainable impacts in terms of innovation, productivity or
employment.” One of the world leaders in the economics of clusters, Gilles
Duranton, titled his 2011 survey “‘California Dreamin’: The Feeble Case
for Cluster Policies.” Yet clusters remain fashionable.
In light of what we have described, the fi rst question is: When are cluster
policies likely to succeed? The answer is that they are most likely to succeed
when there is clear evidence of scale economies and of knowledge creation
together with local knowledge diff usion. Artifi cial intelligence displays these
17. More precisely, productivity is proportional to N 1/ (– 1) where > 1 is the elasticity of substitution between varieties.
Artifi cial Intelligence and International Trade 481
characteristics, though the extent of international knowledge diff usion can-
not be ignored.
The second question is: What policies are likely to work? To answer this
question we turn to the insights of Ajay Agrawal, Director of Rotman’s
Creative Destruction Lab (CDL), and Michael Porter, the business guru
of cluster policies. We start with Agrawal. Agrawal identifi es two problems
with developing AI in the Canadian context. First, there is a shortage of
people with the skills to scale up companies. Agrawal calls these people
1000Xers. Second, the cost of information about a start-up’s quality is so
high that capital markets cannot identify the best and the brightest start-ups.
Agrawal’s CDL addresses both of these problems by linking start-ups with
serial entrepreneurs who can identify a good start-up, tap into 1000Xers for
growth, and pass on valuable information about start-up quality to inves-
tors globally.
Another approach to the question of what policies are likely to work uti-
lizes Porter’s (1990) diamond, which emphasizes four features of clusters:
( a) factor conditions such as universities and an abundant supply of AI sci-
entists, ( b) home- market- demand externalities for AI, ( c) externalities fl owing from suppliers of specialized intermediate inputs into AI such as fi nan-
cial services, and ( d ) a competitive environment. Items b– d involve eff ects that have already been described in our discussion of knowledge spillovers
and lie at the heart of local agglomeration. Item a is a more conventional
economic factor, that is, drive down the price of the key input by subsidizing
its supply. Yet Porter’s research shows that many clusters are driven primarily
by a. That is to say, the single most important policy in practice is simple:
follow Hinton’s advice in training a large number of AI scientists locally.
Our models also suggest two diffi
culties with Hinton’s advice that must
be shored up. First, there is international rather than national knowledge
diff usion due to the fact that, for example, Canadian- trained scientists are
likely to leave Canada for Silicon Valley, China, and other AI hotspots. This
suggests value in programs like those used successfully in Singapore that
require student loans to be repaid if the student does not work in Singapore
for a minimum number of years.
Second, scale in data is a huge problem for a small country like Canada.
To understand appropriate solutions for this, we now turn to the details of
national regulatory environments that aff ect data and the use of AI.
19.4
Behind- the- Border Trade Barriers:
The Domestic Regulatory Environment
Given these models, we next turn to the specifi c regulatory issues that
are likely to impact trade policy. Many of the core trade issues around AI
involve access to data. Data is a key input into AI, and there are a number
of government policies that aff ect data access and data fl ows. To the extent
482 Avi Goldfarb and Daniel Trefl er
these regulations vary across countries, they can advantage some countries’
AI industries. The models above suggest that this advantage can have con-
sequences if there are economies of scale, local externalities, and/or rents.
We highlight fi ve policies in particular. The fi rst three involve data: domes-
tic privacy policy, data localization rules, and access to government data.
The others are development of the regulation of AI application industries
(such as autonomous vehicles) and protection of source code. Privacy policy,
data localization, and source code access have already become signifi cant
trade issues. For example, the TPP addresses all three of these, as do the US
Trade Representative’s NAFTA renegotiation objectives. The US position
is that strong Canadian and Mexican privacy rules, localization require-
ments, and access to foreign source code are all impediments to US exports
of AI- related goods. In other words, the emphasis on trade policy in these
areas is that regulation could be disguised protection that helps domestic
fi rms and hurts foreign fi rms. In the discussion below, we explore the extent
to which this starting assumption is appropriate.
Privacy Regulation. Privacy regulation involves policies that restrict the
collection and use of data. Such regulation diff ers across locations. Privacy
policy has the power to limit or expand the ability of fi rms to use AI eff ec-
tively. Restrictions on the use of data mean restrictions on the ability to use
AI given the data available; however, restrictions on the use of data may also
increase the supply of data available if it leads consumers to trust fi rms that
collect the data. Although the theory is ambiguous, thus far, the empirical
evidence favors the former eff ect on balance. Stricter privacy regulations
reduce the ability of fi rms and nonprofi ts to collect and use data and there-
fore leads to less innovative use of data (Goldfarb and Tucker 2012). Thus,
fi rms in some countries may benefi t from favorable privacy policy.
We believe the most useful analogies for privacy policy in trade relate
to labor and environmental regulations. Such regulations also diff er across
countries for a variety of reasons. They could refl ect diff erences in prefer-
ences across countries, or could be perceived as normal goods that wealthier
countries are willing to pay for but poorer countries are not (Grossman and
Krueger 1995). There is room for reasonable disagreement on how data
might b
e collected or used. Some countries will restrict the information
used in prediction while others will not. For example, for insurance, the data
that can be used varies by state, with diff erent states providing a variety of
restrictions on the use of race, religion, gender, and sexual orientation in
insurance.18 Even with such restrictions, if other variables provide surrogates
for such categories, it is possible that fi rms may be forced to abandon AI
methods entirely for more transparent prediction technologies. In terms of
18. http:// repository.law.umich .edu/ cgi/ viewcontent.cgi?article=1163&context=law_econ _current.
Artifi cial Intelligence and International Trade 483
privacy policy, we think it is useful to take as given that there are diff erences
across countries in their preferences for policies that restrict the collection
and use of data.
Given these diff erences in preferences, what are the implications for trade?
Suppose that the optimal privacy policy for growing an AI industry involves
relatively few restrictions on data. Artifi cial intelligence requires data, and
so the fewer government restrictions on data collection, the more rapidly the
industry grows.19 To the extent that young fi rms tend to grow by focusing
on the domestic market, this will advantage the growth of AI fi rms in some
countries relative to others. Thus, lax privacy policies may help domestic
industry relative to countries with strict policies just as lax labor and envi-
ronmental regulation may help the domestic industry.
This suggests the potential of a “race to the bottom” in privacy policy.
Evidence for such races has been found in enforcement of labor policies
(e.g., Davies and Vadlamannati 2013) and in environmental policies (e.g.,
Beron, Murdoch, and Vijverberg 2003; Fredriksson and Milliment 2002).
There is evidence that privacy regulation does disadvantage jurisdictions
with respect to their advertising- supported software industries. In par-
ticular, Goldfarb and Tucker (2011) examined a change in European privacy
regulation (implemented in 2004) that made it more diffi
cult for European
internet fi rms to collect data about their online customers. This regulatory
change was particularly likely to reduce the eff ectiveness of advertising on