The impact of AI and cross-border data regulation on international trade in digital services

AI is transforming digital services, and mobile apps, which routinely reach users far beyond their home markets, show just how global that transformation has become. This column shows that AI makes these products dramatically more attractive, increasing their user base at home and abroad by more than threefold. But when governments restrict the movement of data across borders, the gains from AI are halved. This creates a genuine policy tension. Rules designed to protect privacy, national security, and democratic institutions can also make it harder to build better products and spread the gains from new technologies.

Digital services are the fastest-growing part of international trade. They already account for about a quarter of world exports and a third of US exports (OECD 2023). Mobile apps are a vivid example. A user in Frankfurt opening Facebook, a listener in Singapore streaming Spotify, or a traveller in Warsaw navigating with Waze is consuming a service developed abroad. When the user and developer are in different countries, the developer is exporting a digital service.

This international reach is not confined to a handful of giant platforms. Figure 1 shows the foreign share of users for apps in our sample. Facebook has about 90% of its users outside the US. Even TikTok, which lies relatively low in the distribution, has almost half of its users outside its home market. Small apps have a foreign-user profile remarkably similar to that of the full sample. Global reach is therefore a basic feature of the app economy, not just a feature of blockbuster firms.

Figure 1 Foreign user shares are high for both large and small apps

Note: The curves show the cumulative distribution of app-level foreign user shares in 2020 through the 60th percentile. Small apps are those at or below the median number of total users.
Source: Authors’ calculations using Sensor Tower data; Sun and Trefler (2026).

We measure an app’s market reach by its monthly active users: the number of people who open the app at least once in a month. This is a central business metric. Around 70% of mobile-app revenue comes from advertising, whose value depends heavily on the number and engagement of users. In the language of the industry, users mean eyeballs and eyeballs mean ad revenue.

AI has become central to this business model because it helps apps personalise content, keep users engaged, and improve products. But AI also depends heavily on data. Foreign users generate data that can be used to train algorithms and personalise the service they receive. This creates a policy conflict. Governments want the gains from digital services, but they are also concerned about privacy, national security, misinformation, and foreign influence. As a result, cross-border data regulation is a dog’s breakfast: some trade agreements promote freer data flows and others restrict data flows through local storage and other requirements. 

These restrictions rose sharply during the same period in which mobile apps and AI were spreading rapidly. Figure 2 shows the increase in restrictions on international data transfers and local storage requirements. 

Figure 2 Restrictions on cross-border data flows rose sharply during the mobile app boom

Note: The series counts data-protection regulations relating to international data transfers and local storage requirements.
Source: Casalini and Lopez Gonzalez (2019), reproduced in Sun and Trefler (2026).

These reflections motivate two simple questions. First, how much does AI actually improve a digital product? Second, because AI feeds on data, what happens to the effectiveness of AI when governments put barriers in the way of those data flows?

Following AI from patents to products

To answer these questions, we need to know how much AI is used in each app. This is difficult because patents are normally linked to firms or broad industries, not to individual products. In a recent study (Sun and Trefler 2026), we combine descriptions of 27,727 apps with the text of 63,679 AI patents owned by their developers. The apps were produced in 59 countries and used in 84 countries.

We use Google’s BERT language model to read the app descriptions and patent texts. BERT converts each text into a numerical representation of its meaning. We then ask how closely the content of each AI patent is related to each app. In total, we calculate 1.8 billion app-patent similarities. Patents whose AI algorithms are closely related to an app receive more weight in our measure of the AI used by that app. In effect, we use an early large language model to trace AI from research inside a firm to the products that consumers actually use. (Much of the research for our paper was conducted before the release of ChatGPT in November 2022.)

There is an important empirical problem: successful apps may be more likely to invest in AI. A simple correlation between AI and users would therefore not tell us the causal effect of AI. Our estimates use an instrumental-variable strategy based on the idea that apps facing similar underlying AI cost conditions should make similar AI investments. We use the AI adopted by a closely matched app to predict AI use in the app we are studying.

AI greatly expands an app’s user base

Our first headline result is stark. An app that goes from using no AI to using a typical (median) level of AI sees a huge increase in users. Our preferred estimate is a 1.26 log-point increase, equivalent to a rise by a factor of 3.5.

This is not simply a Google or Facebook effect. The result remains when we drop the largest apps and when we control for app characteristics, the firm’s other innovative activity, and firm size. Further, the result is driven not by patents per se, but by patents containing AI algorithms that are closely related to the app. This points to the use of relevant AI technology rather than general firm innovativeness or patenting.

Data are AI’s handmaiden

AI algorithms with limited data are of limited value. Data are used both to train models and to personalise predictions, so AI-enabled products continue to improve as users interact with them. Discriminatory cross-border data restrictions may therefore hamper the effectiveness of AI.

To measure the policy environment, we use custom components of the OECD Digital Services Trade Restrictiveness Index for 60 countries. Our measure focuses on rules that discriminate against foreign suppliers. It includes restrictions on cross-border data transfers and local-storage requirements, as well as rules that make it harder for foreign apps to reach local users and generate data.

Our second headline result is equally stark. In countries without discriminatory data restrictions, moving from no AI to the typical level of AI deployment increases users by a factor of 3. At high observed levels of data restrictiveness, the corresponding increase is only 1.5. In this sense, data restrictions cut the impact of AI roughly in half.

Policy implications

Our results do not imply that restrictions on data flows are necessarily undesirable. Privacy, national security, mental health, misinformation, and foreign interference are legitimate concerns. Governments may decide that some restrictions are worth their economic cost.

But there is an economic cost. AI has a very large effect on the reach of mobile apps, and restrictions that limit access to data substantially reduce that effect. Data policy is therefore also trade policy. Discriminatory rules governing cross-border data flows affect which digital products can improve, which firms can expand internationally, and where AI-enabled services can compete.

This tension is likely to become more important over time, not less. With the rise of large language models in everyday life, AI is spreading rapidly through digital services, while governments are becoming more active in regulating the international movement of data. It is high time economists joined this policy debate – not to decide which values should prevail, but to make clear what is gained and what is given up.

Source : VOXeu