by Ajay Agrawal
websites that relied on customer- tracking data. Using a consistent mea-
sure of the eff ectiveness of thousands of online advertising campaigns, the
results showed that European online advertising became about 65 percent
less eff ective after the regulation took eff ect, compared to before the regu-
lation and compared to advertising in other jurisdictions, mainly the United
States. In other words, privacy regulation seemed to reduce the ability of
companies to use data eff ectively. In a diff erent context, Miller and Tucker
(2011) show that state- level privacy restrictions can reduce the quality of
health care. While this evidence does not pertain to AI, just like AI, online
advertising and health care use data as a key input. In other words, the same
forces will likely be at play for privacy regulation that restricts the ability of
AI to operate.
Under strategic trade models, such races to the bottom are likely to matter
if there are rents to be gained from AI. Under endogenous growth models
with local spillovers and various agglomeration models, this could create an
equilibrium in which the AI industry moves to the country with the most lax
policies. Currently, privacy policies are much stricter in Europe than in the
19. Importantly, this is not a statement about the optimal privacy policy from the point of view of a fi rm. If consumers have a preference for privacy, the private sector can provide it even in the absence of regulation. For a richer debate on this point, see Goldfarb and Tucker (2012) and Acquisti, Taylor, and Wagman (2016).
484 Avi Goldfarb and Daniel Trefl er
United States or China.20 Furthermore, there are a number of diff erences in
such policies between the United States and China. This may give the United
States and China an advantage over Europe in this industry.
If stricter privacy policy is likely to hamstring domestic fi rms in favor
of foreign ones, we would expect policy to emphasize avoiding such a race
to the bottom; however, recent trade negotiations have instead focused on
privacy regulation as disguised protection. For example, this argument is
at odds with the current US trade negotiation objectives, which want to
weaken Canadian privacy laws. Based on the existing evidence from other
data- driven industries, we believe this will help the Canadian industry rela-
tive to the US industry in the long run, even if it benefi ts American compa-
nies that already do business in Canada in the short run. In addition, TPP’s
chapter 14 on Electronic Commerce contains provisions that attempt to
limit disguised protection, but contains almost no language that encour-
ages harmonization in privacy policies beyond a request in Article 14.8.5
to “endeavor to exchange information on any such [personal information
protection] mechanisms . . . and explore ways to extend these or other suit-
able arrangements to promote compatibility between them.” The words
“endeavor” and “explore” are what are known in the trade policy literature
as “aspirational” language and generally have no force. The CETA agree-
ment is even more vague with respect to electronic commerce generally.
The electronic commerce section, chapter 16, says little but “recognize the
importance of ” electronic commerce regulation and interoperability and
that “the Parties agree to maintain a dialogue on issues raised by electronic
commerce.”21
It is important to note that this is not a statement about company strategy.
The market may discipline and provide consumer protection with respect
to privacy. Apple, in particular, has emphasized the protection of the per-
sonal information of its customers as it has rolled out AI initiatives, and it
is an open question whether this strategy will pay off in terms of consumer
loyalty and access to better quality, if limited, data.
We also want to emphasize that we do not have a position on the optimal
amount of privacy as enforced by regulation. In fact, we think this is a diffi
-
cult question for economists to answer. Given that the empirical evidence
suggests that privacy regulation, on balance and as implemented thus far,
seems to reduce innovation, the determination of the optimal amount of
privacy should not focus on maximizing innovation (through, as the TPP
20. Canada sits somewhere in the middle. Europe is strict on both data collection and its uses.
Canada’s core restrictions involve use for a purpose diff erent from the collection context. The United States emphasizes contracts, and so as long as the privacy policy is clear, companies can collect and use data as they wish (at least outside of certain regulated industries like health and fi nance).
21. https:// ustr .gov/ sites/ default/ fi les/ TPP- Final- Text- Electronic- Commerce .pdf, http://
www .international.gc .ca/ trade- commerce/ trade- agreements- accords- commerciaux/ agr- acc
/ ceta- aecg/ text- texte/ 16 .aspx?lang=eng.
Artifi cial Intelligence and International Trade 485
emphasizes in article 14.8.1, “the contribution that this [privacy protec-
tion] makes to enhancing consumer confi dence in electronic commerce”).
Instead, it is a balance of the ethical value of (or even right to) privacy and
the innovativeness and growth of the domestic AI industry.
To reiterate, privacy regulation is diff erent from many other regulations
because privacy (perhaps disproportionately) hamstrings domestic fi rms.
Therefore, trade negotiations should not start with the assumption that pri-
vacy regulation is disguised protection. Instead, discussions should start
with the public policy goal of the “social benefi ts of protecting the per-
sonal information of users of electronic commerce” that is also mentioned
in article 14.8.1 of the TPP. Then, if needed, discussions can move to any
particular situation in which a privacy regulation might really be disguised
protection. As we hope is clear from the above discussion, domestic privacy
regulations that restrict how fi rms can collect and use data are unlikely to
be disguised protection. We next turn to two other regulations that might
use privacy as an excuse to favor, rather than hamstring, domestic fi rms.
Data Localization. Data localization rules involve restrictions on the abil-
ity of fi rms to transmit data on domestic users to a foreign country. Such
restrictions are often justifi ed by privacy motivations. Countries may want
data to stay domestic for privacy and (related) national security reasons. In
particular, the argument for data localization emphasizes that governments
want the data of their citizens to be protected by the laws of the domestic
country. Foreign national security agencies should not have access to data
that occurs within a country, and foreign companies should be bound by
the laws of the country where the data were collected. The argument against
such localization (at least in public) is technical: such localization imposes
a signifi cant cost on foreign companies wanting to do business. They need
to establish a presence in every country, and they need to determine a sys-
tem that ensures that the data is not routed internationally (something that
is technically costly, particularly for integrated communications networks
such as within Europe or within North America). US- based companies have
lobbied against such requirements.22
On the technical side, consider two parties, A and B, who reside in the
same country. Internet traffi
c between A and B cannot be confi ned within
national borders without specifi c technical guidance (and some cost to qual-
ity) because the internet may route data indirectly. In addition, data on a
transaction between A and B may be stored on a server located in a diff erent
country. Furthermore, if A and B reside in diff erent countries, then the data
on that transaction will likely be stored in both countries.23
Data localization is an issue for AI because AI requires data. And it often
involves merging diff erent data sources together. The quality of aggregate
22. https:// publicpolicy.googleblog .com/ 2015/ 02/ the- impacts- of-data- localization- on .html.
23. Dobson, Tory, and Trefl er (2017).
486 Avi Goldfarb and Daniel Trefl er
predictions from AI will be lower if the scale of data is limited to within a
country. In other words, localization is a way to restrict the possible scale of
any country in AI, but at the cost of lower quality overall.
Put diff erently, data localization is a privacy policy that could favor
domestic fi rms. Unlike the consumer protection privacy policies highlighted
above, it can favor domestic over foreign fi rms because the foreign- fi rm
AI experts may not have access to the data. The TPP recognizes this and
explicitly restricts it in Article 14.11.3a, which states that the cross- border
transfer of information should not be restricted in a manner that would
constitute “a disguised restriction on trade.”24
Privileged Access to Government Data. Another potential restriction on
trade that might be justifi ed by privacy concerns involves access to govern-
ment data. Governments collect a great deal of data. Such data might be
valuable to training AIs and improving their predictions. Such data include
tax and banking data, education data, and health data. For example, as
the only legal provider of most health care services in Ontario, the Ontario
government has unusually rich data on the health needs, decisions, and out-
comes of 14 million people. If domestic fi rms are given privileged access to
that data, it would create an indirect subsidy to the domestic AI industry.
We think the most useful analogy in the current trade literature is the peren-
nial softwood lumber trade dispute between Canada and the United States.
In the softwood lumber case, most timber in Canada is on government-
owned land, while in the United States, most timber is on privately owned
land. The US complaints allege that Canadian timber is priced too low, and
is therefore a government subsidy to the Canadian lumber industry. While
there have been various agreements over the years, the disagreement has not
been fully resolved. The superfi cial issue is what a fair price should be for
access to government resources. The real issue is whether legitimate regula-
tory diff erences can be argued to convey unfair advantage and therefore
constitute a trade- illegal subsidy.
Government data can be seen similarly. Links between the state and the
corporation vary by country, and this might help some corporations more
than others. What is a fair price for access to the data? Importantly, govern-
ments may not want to give foreign fi rms access to such data for the same
privacy and national security issues that underlie motivations for data local-
24. Related to the issue of data localization is the question of who owns data collected on domestic individuals by foreign individuals or fi rms. For example, consider an American company that uses Peruvians’ cell phones to gather data on agriculture and climate. Who owns the rights to that data? Are the Americans allowed to profi t from that data? Are contracts between the individual actors enough, or is there a need for international laws or norms? The data might not be collected if not for the private companies, but the companies use the data in their own interest rather than in the public interest or in the interest of the Peruvians who provided the data. The recent attempts at a joint venture between Monsanto and John Deere, along with the US Department of Justice antitrust concerns that scuttled the deal, highlight how tangible this issue is.
Artifi cial Intelligence and International Trade 487
ization. Thus, seemingly reasonable diff erences across countries in their data
access policies can end up favoring the domestic industry.
Industrial Regulation. Most international agreements have a section on
competition policy and industrial regulation. This is because regulation can
be a source of unfair comparative advantage or disadvantage. In AI appli-
cations, this list is long. In addition to the points around data and privacy
highlighted above, many applications of AI involve complementary tech-
nologies in which standards might not yet exist and the legal framework
might still be evolving.
For example, in autonomous vehicles, a variety of standards will need to
be developed around vehicle- to-vehicle communication, traffi
c signals, and
many other aspects of automotive design. Most of these standards will be
negotiated by industry players (Simcoe 2012), perhaps with some govern-
ment input. As in other contexts, national champions can try to get their
governments to adopt standards that raise costs for foreign competition.
This leads to the possibility of international standards wars. This is par-
ticularly true of standards that are likely to involve a great deal of govern-
ment input. For example, suppose governments require that the AI behind
autonomous vehicles be suffi
ciently transparent that investigators are able
to determine what caused a crash. Without international standards, diff erent
countries could require information from diff erent sensors, or they could
require access to diff erent aspects of the models and data that underlie the
technology. For companies, ensuring that their AI is compatible with mul-
tiple regulatory regimes in this manner would be expensive. Such domestic
regulations could be a way to favor domestic fi rms. In other words, domestic
technology standards around how AI interacts with the legal regime is a
potential tool for disguised restriction on trade.
The autonomous vehicle legal framework is evolving, with diff erent coun-
tries (and even states within the United States) allowing diff erent degrees of
autonomy on their public roads. Drones are another example where, in the
United States, the Federal Aviation Administration (FAA) strictly regu-
lates American airspace, while China and some other countries have fewer
restrictions. This may have allowed China’s commercial drone industry to
be more advanced than the industry in the United States.25 Thus, regulation
can also impact the rate of innovation and therefore comparative advantage.
Source Code. To the extent that AI may discriminate, governments may
demand information about the algorithms that underlie the AI’s predic-
tions under antidiscrimination laws. More generally with respect to software,
including AI, governments may demand access to
source code for security
reasons, for example, to reduce fraud or to protect national security. Thus,
using consumer protection or national security as an excuse, governments
25. https:// www .forbes .com/ sites/ sarahsu/ 2017/ 04/ 13/ in- china- drone- delivery- promises
- to-boost- consumption- especially- in-rural- areas/ #47774daf68fe.
488 Avi Goldfarb and Daniel Trefl er
could reduce the ability of foreign fi rms to maintain trade secrets. Further-
more, cyber espionage of such trade secrets may be widespread, but that is
beyond the scope of this chapter.26 Broadly, this issue has been recognized in
the TPP negotiations, with Article 14.17 emphasizing that access to source
code cannot be required unless that source code underlies critical infrastruc-
ture or unless the source code is needed to obey other domestic regulations
that are not disguised restrictions on trade.
Other policies that might aff ect the size of domestic AI industries include
intellectual property, antitrust, R&D subsidies, and national security. If AI
is the next important strategic industry, then all of the standard questions
arise with respect to trade policies in these industries. We do not discuss these
in detail because we think the trade- specifi c issues with respect to these poli-
cies are not distinct to AI, but are captured more generally by the discussion
of innovation and trade. The main point for these other aspects of domestic
policy with respect to AI and trade is that there are economies of scale in AI
at the fi rm level. Furthermore, we expect some of the externalities from the
AI industry to remain local.
19.5 AI and International Macroeconomics
Before concluding, it is important to recognize that AI will have implica-
tions for international macroeconomics. For example, suppose that China
does succeed in building a large AI industry. This will likely increase its trade
surplus with the rest of the world, particularly in services. Furthermore,
suppose that China manages to control wage infl ation through promoting
migration from rural to urban areas, and by relaxing the one- child policy.
Then, this is likely to put upward pressure on the renminbi (RMB) and
downward pressure on the dollar.
This will have implications for US labor markets. At the low end of the