Introduction

Economic espionage represents a substantial threat to US firms and industries. As China has grown economically, it has pursued various strategies to acquire knowledge and technology from US firms, with economic espionage being one of the most direct methods. From high-profile cyber intrusions targeting semiconductor companies to insider theft at pharmaceutical firms, espionage incidents have become increasingly common. The US government has responded with heightened enforcement through the Economic Espionage Act and the China Initiative, reflecting concerns about the economic damages from these incidents . Despite this policy attention, systematic empirical evidence on the economic consequences of espionage remains limited. This makes it important to understand: what are the economic harms from espionage to targeted firms and US industry, and how do firms respond to being targeted?

To answer these questions, we compile a comprehensive dataset on publicly disclosed espionage incidents based on prosecutions under the Economic Espionage Act since 1996 and publicly disclosed cyber intrusions. China is the beneficiary country in 80% of cases in our data. We link this case-level data to firms, which allows us to use an event-study framework to estimate the impact of espionage on firm outcomes. We find that espionage has large negative effects on targeted firms. Being successfully targeted for espionage reduces firm revenue and R&D spending by 40% after 5 years, with effects persisting for 10 years. These results are robust to alternative identification strategies, and they hold for a variety of other firm outcomes including total assets and intangible assets. Importantly, we find no effect on any of these outcomes when espionage attempts are unsuccessful (i.e. when no knowledge was transferred). This placebo test provides strong evidence that our estimated effects are not driven by violations of the parallel trends assumption, but rather reflect the causal impact of successful espionage on firms.

To probe the mechanisms behind these effects, we examine heterogeneity across different types of incidents. We find that the revenue losses from espionage are concentrated in cases where firms claim high economic losses, and in industries where industry-wide R&D spending is relatively low. The first result suggests that espionage is most harmful when it targets key technologies that are difficult to replace. The second result suggests that espionage is most harmful in industries where the key technologies are established (and thus spending on developing new technology is low). Together, these heterogeneity analyses paint a picture whereby espionage makes a firm’s valuable technology obsolete.

Does espionage simply affect firms who are targeted, or does it affect US industry as a whole? We map espionage incidents to industries based on the primary sales sector of targeted firms, and subsequently link incidents to international trade flows in the affected industries. On a 10 year horizon, we find that American exports in targeted sectors decline by 40%. These export results are more uncertain than the firm-level results, because the industry-level placebo test is less clean: Chinese exports rise even in sectors where espionage was unsuccessful, implying that the appropriate counterfactual for China involves a rising baseline. With this caveat, we find no evidence that Chinese exports rise through increased international market share, suggesting that the harms to the US are not offset by observable gains to China through this channel. Global exports in targeted sectors decline by 10%, suggesting these are net losses, not simply a US-to-China redistribution. These results demonstrate that the economic damages from espionage extend beyond individual firms to have measurable aggregate effects on US industry. Moreover, they suggest that Chinese firms are not able to capture the full advantage created by espionage through international markets.

Having established the substantial economic harms from espionage, we turn to a natural follow-up question: how do targeted firms manage their innovation in response to espionage? We study firm responses across four margins: patenting behavior, employment of foreign scientists, participation in international markets, and corporate speech. We link firms to patents, OPT visa records from ICE, Compustat segment data, and earnings call transcripts to test for responses along each of these dimensions.

Across all of these outcomes, we find no evidence that firms restrict their innovation management in response to espionage. Espionage does not affect the patent-to-R&D ratio, international patent collaborations, or citations to targeted firms’ patents. Our primary evidence on employment uses OPT visa records, which show that targeted firms do not reduce their hiring of Chinese advanced-degree STEM workers after espionage. Complementing this, Revelio Labs LinkedIn data shows that employment declines after espionage are proportionally similar for Asian and non-Asian workers overall. We do find a marginally significant larger decline among Asian scientists specifically; however, we find no decline in firm-level diversity and inclusion scores, suggesting this pattern reflects Asian scientists having better outside options rather than firm-side discrimination. We also find no evidence that firms exit Chinese or other geographical markets after espionage. Finally, an analysis of earnings call transcripts reveals that firms do not discuss espionage (or strategic responses to it) publicly even after it becomes known, with only two stylistic differences out of over 16,000 tested features of corporate speech.

Together, our results establish that economic espionage causes substantial economic harm to both targeted firms and US industry more broadly, but find little evidence that firms respond by restricting knowledge flows in ways that could further harm innovation. In short, we document clear economic damages from espionage, but do not find that firms change their management of innovation and knowledge sharing in response.

Related Literature

This paper relates to four strands of the literature. First, we contribute to research on the value of trade secrets. Our paper is closely connected to two recent papers studying the effects of trade secret theft and espionage on firms. examine how corporate innovation responds to the theft of trade secrets, while study the value of trade secrets as imputed from espionage incidents. Relative to these papers, we make three contributions. First, each paper’s analysis focuses on a single outcome in targeted firms (innovation and market capitalization, respectively), while we examine a much broader set of firm outcomes including revenue, R&D, assets, patenting behavior, employment discrimination, and corporate speech. Second, our analysis of unsuccessful attempts at espionage provides a stronger argument for causality. Third, we analyze the industry-level effects of espionage using international trade data, allowing us to characterize the net effects of espionage on US industry and show that the damages extend beyond redistribution between firms within the US. Fourth, our dataset of espionage incidents is more comprehensive in scope than the datasets of these two papers (covering 164 incidents, compared to 52 incidents and 72 incidents respectively in these two papers). This makes our analysis more generalizable, despite the difficulty of observing espionage empirically.

Second, we contribute to research on geopolitical tensions and international economic relationships. The growing literature on geopolitical tensions has primarily focused on trade and investment relationships between countries , with knowledge and technology relationships being more difficult to study. Some papers within this framework study academic scientific relationships between the US and China , or industrial espionage during the Cold War . Our contribution is to provide systematic evidence on how economic espionage affects firms and industries today, and to test whether it induces restrictions on innovation that could have broader economic consequences.

Third, we speak to the literature on creative destruction and business dynamism. The classic Schumpeterian model of firm dynamics and growth has been a guiding framework for understanding how firms creatively destroy one another through innovation. A central question in these models is how competition affects innovation incentives, with theory predicting ambiguous effects: competition can spur innovation by increasing the returns to escaping competition (the “escape competition” effect), but can also reduce innovation by decreasing the rents available to innovators (the “Schumpeterian” effect). Building on these models, recent work on US business dynamism highlights “knowledge diffusion”—the ease with which ideas spread from a leader firm to other firms in an industry—as the central factor explaining the decline in US business dynamism . Our setting provides an unusual empirical window into this process. Economic espionage represents an exogenous shock to the competitive pressure faced by industry leaders, as it allows competitor firms (particularly Chinese firms) to acquire the technologies that gave leaders their competitive advantage. Our finding that espionage reduces R&D spending by targeted firms provides evidence that increased competition reduces innovation, consistent with the Schumpeterian channel dominating in our setting.

Finally, we speak to the small literature on the Economic Espionage Act and its consequences. and analyze EEA cases through the lens of whether they are discriminatory towards ethnically Chinese defendants. Relative to this literature, we use the EEA primarily as a window into economic espionage and its economic effects on firms and industries. We also contribute by testing whether counter-espionage concerns lead to discrimination in employment, finding that firms do not appear to become more discriminatory after espionage incidents.

The rest of this paper is organized as follows. Section 2 describes the context of economic espionage and the US-China relationship. Section 3 describes the construction of our database of espionage incidents and other data sources used, and provides descriptive statistics on our data. Section 4 examines the economic damages from espionage on targeted firms and industries. Section 5 analyzes whether firms respond to espionage with innovation restrictions. Section 6 concludes.

Context

From Roman monks stealing the secret of silk making from China in the 6th century, to George Washington and Alexander Hamilton endorsing a program to target the British textile industry to gain “secrets of extraordinary value,” economic espionage has long been recognized as a tool of statecraft. In the 19th century, the US chemical industry grew to the technological frontier through hiring German chemists who brought trade secrets . After World War II, the US recruited hundreds of German scientists through Operation Paperclip, recognizing their importance in the oncoming Cold War. During the height of the Cold War, East Germany kept pace with West Germany in its industrial development through espionage against West German industry . In short, countries catching up to the technology frontier through economic espionage has a long history.

Today, economic espionage is believed to be widespread. For example, in a report uncovering a series of coordinated breaches of over fifty global companies, the security firm McAfee makes the staggering claim (emphasis not added):

I am convinced that every company in every conceivable industry with significant size and valuable intellectual property and trade secrets has been compromised (or will be shortly), with the great majority of the victims rarely discovering the intrusion or its impact. In fact, I divide the entire set of Fortune Global 2,000 firms into two categories: those that know they’ve been compromised and those that don’t yet know.

The Economic Espionage Act

The United States’ approach to economic espionage underwent a fundamental shift with the passage of the Economic Espionage Act (EEA) in 1996. Prior to the EEA, trade secret theft was primarily addressed through civil litigation, leaving prosecutors with limited tools to combat state-sponsored economic espionage. The Act established two key criminal offenses: theft of trade secrets to benefit foreign entities (Section 1831) and domestic trade secret theft (Section 1832). Penalties under Section 1831 are particularly severe, reflecting Congress’s concern about foreign economic espionage, with individual defendants facing up to 15 years imprisonment and fines up to $5 million.

The EEA’s focus has evolved significantly over time. Its initial motivation was based on concerns about industrial espionage by France, who had made it clear that they wanted to develop a computer industry to compete with the US’s computer industry, and were suspected of using industrial espionage to aid that process . However, in the 2000s, the focus shifted toward China, reflecting broader changes in geopolitical tensions and technological competition. From 1997-2008, defendants of Chinese descent represented 17% of EEA prosecutions. This proportion increased dramatically to 52% during 2009-2015 . This shift coincided with China’s rapid technological advancement, and increasingly explicit policies aimed at acquiring foreign technology and expertise.

US-China Rivalry and Espionage

As China has grown, it has pursued ways to gain knowledge from the US. These efforts are exemplified by the Thousand Talents Program, launched by China in 2008. The Thousand Talents Program represents one of China’s most significant efforts to acquire foreign technology and expertise. It aimed to recruit leading international scientists, researchers, and entrepreneurs to work in China, with a particular focus on those with expertise in strategic technologies and access to intellectual property at major Western institutions. By 2018, over 7,000 individuals had participated, including both Chinese nationals working abroad and foreign scientists. Although research collaboration in most forms is legitimate, the program has also been used to facilitate unauthorized technology transfer, including several incidents in our sample. The Thousand Talents Program evolved significantly after receiving heightened scrutiny. China has stopped publicizing information about participants and rebranded various components of the program.

In the context of countering the Thousand Talents Program, the US Department of Justice launched the China Initiative in 2018. The China Initiative marked an aggressive shift in US law enforcement’s approach to Chinese economic espionage. The initiative prioritized investigating and prosecuting cases involving technology theft benefiting China, with a particular focus on academic and research institutions. It represented the first time the DOJ focused its national security resources explicitly on a single country, and made it clear that the US was going to invest in countering China’s talent program.

The China Initiative expanded traditional counterintelligence work in several ways. It increased scrutiny of research collaborations, requiring academics to disclose Chinese funding sources and affiliations. For example, the FBI maintains a page about China’s talent plans . This page proactively warns companies and academic researchers about the risk of espionage from their students and employees (emphasis added):

Talent plans can sometimes foster legitimate sharing and collaboration as part of an appropriate business arrangement or research exchange, but this is not the norm. Instead, talent plans usually involve undisclosed and illegal transfers of information, technology, or intellectual property that are one-way and detrimental to U.S. institutions. Your students and/or employees could be talent plan participants.

The China Initiative also broadened prosecution strategies, using tools like grant fraud charges and failures to disclose foreign ties rather than relying solely on evidence of direct technology theft. This enabled prosecutions even without direct evidence of unauthorized technology transfer.

By 2021, the DOJ had brought over 70 cases under the initiative, charging both Chinese nationals and American citizens, particularly those of Chinese descent. However, the initiative faced mounting criticism over its implementation. Civil rights organizations argued it disproportionately targeted Asian-American scientists—one study found that 88% of defendants charged with grant fraud were of Chinese heritage . Critics argued that the initiative’s broad scope created a chilling effect on legitimate scientific collaboration and contributed to racial profiling in academia and industry science. In January 2022, the New York Times documented a high-profile dismissal of all charges against MIT professor Gang Chen , bringing the criticism of the China Initiative to its peak. Just one month later, the DOJ ended the China Initiative, replacing it with a broader “Strategy for Countering Nation-State Threats”, which eliminated China-specific branding, though it has maintained a focus on Chinese espionage.

Vignettes of Economic Espionage

It is useful to characterize what incidents of economic espionage actually look like. An emblematic case comes from Orbit Irrigation, an American company that manufactures sprinkling and irrigation systems. Janice Kuang Capener was in charge of operations at an Orbit plant in China from 2003 to 2009. She provided proprietary information about Orbit’s pricing strategy to Zhejiang Hongchen Irrigation Equipment, a Chinese competitor to Orbit, which they used to undercut Orbit’s position in the market . This case demonstrates that economic espionage is not always about the theft of technology, but is sometimes about economic information or market research that can also be important to businesses.

Another example is the case of Valspar Corp, an American paint company. David Lee—a technical director at Valspar at the time—illegally downloaded trade secrets from Valspar and Huarun (a Chinese subsidiary of Valspar) to an external thumb drive, with the intention of giving these secrets to the Shanghai office of Nippon Paint (a Japanese paint company) where he had accepted a vice-president position. However, he was arrested before he could actually deliver the information to Shanghai . This mechanism of espionage—an individual downloading information illegally with the benefit of financial reward and employment in China—is representative of many cases. This case also exemplifies an “unsuccessful” case of espionage—where no information was actually leaked from Valspar. This kind of case will later be important for assessing the credibility of our results.

Data and Descriptive Statistics

To examine the impacts of economic espionage on victim firms and their responses, we construct a novel dataset that combines information on espionage incidents with detailed firm-level outcomes. This section outlines the key data sources and the construction of our analysis sample.

Espionage Data

Our primary data on economic espionage come from court filings, Department of Justice (DoJ) announcements, and reports from relevant think tanks. This dataset compiles all publicly documented cases brought under the Economic Espionage Act since its enactment in 1996, as well as other documented instances of foreign economic espionage against US firms since 1990. For the purposes of this study, we define economic espionage as theft of a firm’s proprietary economic information—these are most often technological trade secrets, but may also include confidential information such as business development or pricing strategies. We do not include cases of explicit forced technology transfer, such as through China’s quid pro quo policy . We also omit cases where the only data stolen is customer information, such as data breaches of emails or passwords, since this kind of information is not central to a firm’s economic advantage over its competitors.⁴

Sample Construction

To build a comprehensive database of incidents of economic espionage against the United States, we make use of the following sources:

1. First, we query CourtListener, a non-profit legal research database that contains legal data from a variety of sources, including the commonly used Public Access to Court Electronic Records (PACER) data.⁵

2. Second, we search through all Department of Justice press releases about economic espionage indictments,⁶ alongside relevant publications such as the DOJ’s Pro IP Act Annual Reports.⁷

3. Third, we complement these primary sources with compilations of espionage cases created by other third parties. These come from a wide range of actors, specifically the Cato Institute , the Center for Strategic and International Studies , investigative journalists , legal scholars and concerned citizens .

4. Finally, we conduct our own search of news articles, corporate press releases, and cybersecurity firm announcements.⁸

All of these sources are necessary to develop a comprehensive picture of economic espionage, because not all incidents are pursued or prosecuted by the US government (and so may not appear on the DOJ websites or court filings). We manually confirm whether each economic espionage incident occurred or not, and collect the following information for each incident: date espionage began, date espionage ended (if distinct), earliest date espionage was reported or announced to the public, whether the suspect successfully obtained the proprietary information, the suspect’s country, whether this was a cybersecurity incident or not, alongside the victim companies targeted, and the suspect firm (when available).⁹

We manually code whether an espionage incident was “successful”, meaning that the attempt to steal trade secrets was successful and the information was delivered to some beneficiary, or “unsuccessful”, meaning that the attempt was not successful. This distinction is generally unambiguous from the descriptions of cases in press releases or court documents.

Descriptive Statistics

We uncover 164 cases of attempted economic espionage in our primary sample. Of these, 126 are successful (77%, cases where the suspect successfully delivered the proprietary information to the intended target), and 129 of these are attributed to China (79%). We show a timeline of espionage incidents in Figure 1.

In Table [t:targeted_industries], we provide a list of industries (SIC codes) targeted more than once by espionage, as well the number of times these industries have been targeted. The most targeted industry is semiconductors. Frequently targeted industries, including plastics, pharmaceuticals, and aircraft manufacturing, tend to be knowledge-heavy in nature, as well as reliant on trade secrets to protect their competitive advantage. They are also more likely to have military significance; in Figure 2 we plot the share of industries that are “dual-use” by whether they are targeted by espionage or not. We find that 88% of SIC codes targeted in our data are dual-use, compared to 47% of SIC codes in general. Thus, targeted SIC codes are twice as likely to be dual-use as the average SIC code, consistent with strategic motivations for espionage.

More than three-quarters of our incidents take place in a repeatedly targeted industry. In Figure 3, we plot the rank of victim firm revenues (before espionage) relative to their industry. We find that targeted firms tend to be leaders in their industry: in over half of cases, they are one of the three largest (public) firms in their 4-digit SIC code. However, Figure  8 shows that the same pattern does not hold when examining firms relative to their 3-digit SIC code: targeted firms are spread much more evenly through the firm rank distribution in their 3-digit SIC code. This suggests that the targeting of firms occurs within a 4-digit SIC code rather than the 3-digit code. Thus, the 4-digit SIC code is likely to provide the best definition of a "targeted industry", and we use it as our industry definition in our analysis.

Analysis of Suspects

While the dataset identifies victim firms comprehensively, limited information exists on the perpetrators of espionage. Only 27 cases involve suspect companies that are publicly traded. We are unable to identify a suspect firm in the majority of cases for a variety of reasons: many individuals who conduct espionage incidents intend to found their own start-up or give the information to a small enterprise for which we cannot find data. In other cases, espionage is attributed to foreign universities or government agencies (such as the Chinese Ministry of State Security), making it unclear which firms precisely benefit from espionage. As a result, even though the effect of espionage on suspect firms is an important question, our data is too noisy to answer it. Thus, our analysis of the effects of espionage on China uses international trade data from UN COMTRADE, which allows us to examine the aggregate spillover effects of economic espionage across countries.

Sample Selection

We acknowledge a major limitation of our study: we can only observe espionage in the selected set of cases that are detected and reported. This set may be very different from the typical espionage case, which may still remain undetected, or only detected by a security firm and never reported to the world. Thus, we cannot claim to offer a full accounting of the welfare costs of economic espionage, or an understanding of the average espionage case (which may be very different from what we observe).

All of our results should be interpreted in the context of this selected sample. For our research question, economic espionage incidents are best understood as a window into how firms react and change their openness to knowledge diffusion in response to economic espionage information shocks. In that regard, selecting the most high-profile and important cases is an advantage, rather than a disadvantage, because these cases are more likely to drive changes in US-China knowledge diffusion than unknown cases.¹⁰

Firm-Level Data

Out of 164 cases of economic espionage, we are able to identify a victim firm in 130 incidents. There are 142 unique firms that we identify as targeted by espionage (some incidents target more than one firm). Of these 142 firms, 108 are ever publicly listed (or are subsidiaries of listed firms), allowing us to collect detailed financial information on them. Of these public firms, 75% of cases are successful incidents, and 83% are attributed to China.

Firm Financial Data

To analyze the economic consequences of espionage, we draw on firm-level fundamentals from Compustat (for American firms) and Worldscope (for international firms), standard sources for financial and operational data on publicly traded firms. This provides us with each publicly traded firm’s revenue, R&D expenditures, total assets, intangible assets, gross margins, and net profit margins. We collect data on all firms targeted by espionage, as well as all firms sharing a SIC or SIC-3 code with these firms. This leads to a total of 23,830 firms in our main sample. We use data from 1995 to 2024 on an annual basis. We also supplement this data with Compustat’s segments data to decompose a firm’s sales across locations and industries.

For firm-level stock market returns, we rely on CRSP (for American firms) and Datastream (for international firms). We additionally make use of the standard Fama-French factors made available through WRDS to estimate normal stock returns for all firms.

Firm Employee Data

We examine how firms respond to espionage through their employment composition by using data from Revelio Labs. This dataset, derived from LinkedIn, provides granular information on a firm’s workforce composition. Unlike the data sources above, every single victim firm that we identify has a LinkedIn page, allowing us to analyze outcomes for non-public firms as well. Revelio predicts the race of each employee based on their name, classifying at the level of “Asian" rather than “Chinese."¹¹ We primarily focus on employment of Asian versus non-Asian employees in the US.

Revelio classifies all job titles into 1500 different categories—we use these categories to identify whether an employee does scientific/R&D work in their role. We manually classify these categories based on whether they are likely to conduct research and development work (examples of scientists include “Clinical research”, “Process development engineer”, or “Scientist”).¹² 8.7% of all employees in our sample are classified as scientists. Finally, from Revelio, we also collect employment counts across countries to investigate the geography of employment. This data allows us to understand how firms respond to espionage through their employment of scientists, as well as to detect racial heterogeneity that could reflect China-specific concerns.

Firm Innovation and Knowledge Flows

We also link firm-level patent and innovation outcomes using data from the United States Patent and Trademark Office (USPTO). This database provides comprehensive information on firm-level patenting activity, including location of collaborators, and those who cite these patents. Even in our set of publicly listed firms in high-technology industries, patenting is relatively rare, with 75% of firm-years recording no patents at all. Thus, we focus our analysis of patents on the extensive margin; whether firms patent at all in a given year, and conditional on patenting, who they patent with. From patent records, we pull inventor information, to understand how firms to respond to espionage through the characteristics of inventors on their patents. In particular, we use the address of inventors to characterize where they live, to understand whether espionage leads firms to reduce patenting with inventors living in China or outside the US in general. Finally, we use patent citation information to measure whether patents by targeted firms are any less likely to cite Chinese patents, or to be cited by Chinese patents. Together, these measures give us a rich understanding of how firms endogenously change their knowledge flows with China and the rest of the world in response to espionage.

Direct Effects of Espionage

This section focuses on estimating the direct effect of economic espionage on firm revenues and innovation behaviors. This is a first stage of analysis before we can explore any second-order effects of espionage. For example, if espionage had no effects on firm revenues, we might not expect knowledge diffusion to change as a result of espionage. Thus, in this section we investigate the direct effects of economic espionage on targeted firms.

To estimate the causal effects of economic espionage on firm outcomes, we employ a dynamic difference-in-differences framework. This approach leverages variation in the timing of espionage incidents across firms to identify treatment effects, as well as which firms are ever targeted, while accounting for firm-specific and sectoral trends.

Baseline Specification

Our core empirical specification for understanding how espionage affects firm revenue and R&D is given by the following dynamic two-way fixed effects (TWFE) regression:

$$\begin{equation} y_{i,t} = \sum_{k = -4}^{10} \gamma_k D_{i,t+k} + \alpha_i + \alpha_{j(i),t} + \varepsilon_{i,t} \label{eq:twfe} \end{equation}$$ where i denotes a firm, t a year, and j(i) the industry firm i belongs to. We interpret the timing of treatment to be when espionage started against a targeted firm. D_{i, t + k} = 1{t − E_i = k}, where E_i denotes the year espionage began against firm i, so that each firm-year observation is assigned to exactly one relative-time bin k (with one bin omitted for normalization). The coefficients γ_k trace the evolution of outcomes before and after espionage begins, allowing us to examine both the pre-trend dynamics and post-treatment effects on various firm outcomes.

An important feature of this specification compared to a standard event study is that our year fixed effects are actually industry-year fixed effects. In other words, we are comparing targeted firms to untargeted firms within the same industry. This feature is important because industries could be selected in whether they are targeted for espionage or not, based on the trends of firms in that industry. Thus, we want to control for industry-specific trajectories when determining whether espionage affects a firm.

To facilitate interpretation of our results and extend the time horizon over which we can measure effects, we interpret espionage as an absorbing treatment. That is, for any firm targeted multiple times by espionage, we use the first time it experiences an incident as the treatment date, and drop later espionage events. Standard errors are clustered at the firm level. The key identifying assumption is that in the absence of espionage, the outcomes of targeted firms would have evolved similarly to those of untargeted firms within the same industry.

We estimate effects on two main firm outcomes: revenue and R&D expenditures. We use revenue as a summary statistic for how much firms are affected by the loss of the competitive advantage they previously had from having proprietary knowledge/technology. However, the welfare effects of espionage are likely to be dominated by how it affects firm innovative activity.¹³ As a result, we use R&D expenditures to capture this dimension of the effects of espionage.

Figure 4, Panel A, plots the dynamic treatment effects on firm revenue and R&D expenditures for firms hit by an espionage incident. We find that firm revenues decline by 40% within 5 years of espionage beginning, with effects persisting for the 8 years of our window. Espionage also appears to reduce firm R&D spending by a similar magnitude, but this effect is more gradual and noisier. These results show that firms substantially shrink (or fail to grow) as a result of economic espionage, and the decline in R&D expenditures suggests that espionage may have important consequences for a firm’s innovative capacity. We find no evidence of pre-trends in the event study.

Placebo Test: Unsuccessful Cases

In addition to the standard pre-trends test, our setting allows a unique approach to testing the validity of our identifying assumption. The key feature of our setting is that we not only observe successful espionage cases, we also observe cases of espionage that failed—i.e. no knowledge was transferred out of the targeted firm. This usually happens because the attempt to steal knowledge was detected and stopped before the theft could occur successfully. This is illustrated by the case of Valspar and David Yen Lee discussed in Section 3. There are 38 such cases in our dataset.

Then under the assumption that successfully-targeted and unsuccessfully-targeted firms would have had parallel trends in the absence of espionage, the sample of unsuccessful cases actually allows us to test whether our parallel trends assumption holds in the post-period. Formally, let S_i = I[espionage on i is successful]. Then under the assumption that 𝔼[y_i, t(0)|D_i, t = 1, S_i = 1] = 𝔼[y_i, t(0)|D_i, t = 1, S_i = 0] estimating Equation [eq:twfe] in the set of unsuccessful cases (S_i = 0) provides a direct test of the parallel trends assumption. This set of cases is excluded from our main estimation sample. If our empirical design is valid, we expect to find no significant effects in this placebo sample (γ_k = 0 for k > 0), as no actual knowledge theft occurred. Any significant effect detected in this placebo sample would be evidence of a violation in parallel trends. Conversely, a null effect in this placebo sample provides evidence for the parallel trends assumption—stronger support than is normally available in difference-in-difference research designs.

In Figure 4 Panel B, we plot results for the placebo test using unsuccessful cases of espionage, and find null results, also supporting a causal interpretation of our main estimate. Failed attempts to steal information do not harm the targeted firms.

To verify that this null effect in the placebo test is not simply noise caused by a small sample size, we estimate a triple-difference specification in Table [t:triplediff] that allows us to specifically test whether the treatment effect differs between the two groups (successful firms and unsuccessful firms). We find that even with the small size of our placebo sample, we can reject that the placebo effect on revenue is the same as the effect on successful firms—providing evidence that our effects are not simply driven by pre-trends. However, R&D is a noisier outcome, and the effect we show in Figure 4 only becomes large many years after the espionage. This is why the triple difference coefficient, which includes the years just after espionage, is smaller and thus can’t be separated from the placebo.

The validity of this placebo test rests on Assumption [eq:placebo]: that successfully and unsuccessfully targeted firms would have had parallel trends absent espionage. We acknowledge this assumption is not trivially satisfied—in principle, firms with weaker internal security may be both more likely to have espionage succeed and more likely to decline for unrelated reasons. Nonetheless, this assumption is substantially weaker than the standard parallel trends assumption required for a simple difference-in-differences design, which requires every non-targeted firm to be a valid counterfactual for every targeted firm. Here, we only require comparability between two subgroups that are both targeted for espionage—with the only distinction being whether the attempt succeeded or failed, a determination that often hinges on idiosyncratic factors such as whether an employee was caught before transmitting stolen information. To further validate this assumption, we compare pre-espionage observables between successfully and unsuccessfully targeted firms. We present a balance table in Table [t:balance_succ], finding no statistically significant differences along any characteristic between firms successfully and unsuccessfully targeted by espionage.

Mechanisms

To understand the mechanism by which espionage affects firms, we estimate heterogeneous effects of espionage on firms and plot the results in Figure 5.

First, we separate cases based on whether the damages claimed by the victim firm are above or below the median claimed damages, and estimate Equation [eq:twfe] separately for these two groups. Panels A.1 and A.2 show that our estimated revenue effects are concentrated entirely in cases with above-median claimed damages. Claimed damages are a legal construct that firms calculate based on the self-assessed market view of their stolen technology. Thus, cases with high claimed damages are cases where the stolen technology was especially critical to the firm’s competitive edge, and espionage may represent the loss of this competitive edge.

Second, we separate cases based on whether the targeted industry’s baseline R&D spending is above or below the median for targeted industries. Panel B.1 shows that the revenue damages from espionage are most pronounced in industries with below-median R&D spending. Although potentially counterintuitive, our preferred interpretation is that espionage is most damaging when a firm is in an industry with relatively few key technologies, compared to in a more fast-moving sector where technologies become obsolete quickly and/or the marginal value of each technology is lower.¹⁴

Third, we separate cases based on how concentrated the targeted industry is (as measured by HHI). Panel C.1 shows that the revenue effects of espionage are common for both low-HHI and high-HHI industries, but Panel C.2 shows that espionage reduces R&D only in low-HHI industries. This heterogeneity suggests that in more competitive industries, falling revenue necessitates reduced R&D spending—but in concentrated industries, targeted firms (who are usually the industry leaders, as shown by Figure 3) still have enough rents to sustain R&D spending.

Together, these heterogeneity analyses paint a picture whereby espionage makes a firm’s valuable technology obsolete. Thus, espionage is less harmful when these technologies are less valuable or when the industry’s high rate of R&D spending means the natural obsolescence rate of technology is higher.

Robustness Checks and Additional Analyses

We briefly report additional robustness checks, which are described in more detail in 8. First, while our main specification is a dynamic two-way fixed effects estimator, we show that our results are robust to a variety of alternative event-study estimators (Figure 9). Second, in addition to firm revenue and R&D expenditures, we show that espionage also reduces a firm’s total assets and intangible assets (Figure 10), and that both domestic sales and exports to China decline (Figure 12). Third, we relax the assumption that firms in the targeted industry are untreated, allowing for spillovers to those firms (Figure 11), and find no evidence of meaningful positive or negative spillovers to competitor firms within the same SIC-4 code. Fourth, we look at how firms’ stock returns are affected by both the onset and announcement of espionage, finding that both events lead to persistent declines in returns, consistent with losses from both real factors and the market’s perception of these incidents (Figure 13). Finally, we show that results are unlikely to be driven by mean-reversion of leader firms (Figure 14) or by selection of targeted industries (Table [t:comp_leader_dyn]).

Effects on an Aggregate: Exports

From a policy standpoint, what matters more than effects on targeted firms is the aggregate effect of espionage. If targeted firms lose revenue and conduct less R&D, but non-targeted firms gain revenue and conduct more R&D, the net result could be zero, simply reflecting a reallocation between firms. Policymakers would be much less concerned about espionage in that case, especially if the reallocation happened within a country.

To understand these aggregate effects, we test how espionage incidents affect exports in both the US and China. We use trade data to capture industry-level effects for three reasons. First, it covers all firms operating in each country, including non-publicly listed firms that would not appear in Compustat. Second, in Ricardian models of trade, exports reveal a country’s productivity in a sector, so changes in exports can be interpreted as the best proxy for changes in industry productivity. Finally and most importantly, exports are the cleanest way by which we are able to estimate how espionage affects China. For the rest of the analysis, we cannot reliably link incidents to suspect firms, and even if we could, the benefits of espionage could be diffused across Chinese industry. Therefore, looking at how Chinese exports evolve in targeted sectors is key to understanding the other side of the equation.¹⁵

We map SIC codes to HS6 codes using the 2018 revision of the concordance from , and define treatment at the industry level as the first period in which an industry is ever targeted by an espionage incident. In other words, we keep only the first time any firm in the industry was targeted. In this specification, our never-treated HS6 codes are “nearby” codes—those that share an HS4 code with a targeted HS6 code, but are never themselves targeted for espionage. Our baseline specification is, for the outcome log(exports by country i in sector s at year t):

$$\begin{equation} y_{i,s,t} = \sum_{k=-4}^{10} \left(\gamma_k D_{s,t+k} + \beta_k D_{s,t+k} \mathbb{I}[i \text{ is US}] + \delta_k D_{s,t+k} \mathbb{I}[i \text{ is CN}]\right) + \alpha_{i,s} + \alpha_{i,t} + \epsilon_{i,s,t} \label{eq:trade_ddd} \end{equation}$$

Regressions are at the exporter-HS6-year level, and include exporter-HS6 and exporter-year fixed effects.¹⁶ Apart from the direct treatment effect, we also conduct triple differences, interacting time to treatment with the exporter being the US or China. For interpretation of coefficients, in this section, we restrict cases to only the 129 where the suspect country is China.

We plot results in Panel A of Figure 6. We find that ten years after espionage occurs in a sector, exports by the US fall by 40%. Puzzlingly, exports from China also fall in the short term before recovering in the long term, suggesting that China does not appear to gain through international markets in a way that offsets US losses.

These effects are surprisingly large. Are they due to selection of which industries are targeted for espionage? We test this with the same placebo test as before: under Assumption [eq:placebo], we can test whether these effects are due to espionage by estimating Equation [eq:trade_ddd] in the sample of unsuccessful cases. Panel B of Figure 6 reports the result of this placebo test. Here, the placebo test is not as clean as before. We see no decline in exports from the US or the rest of the world, suggesting that the effects on the US and the rest of the world from Figure 6 are truly causal. However, we do see Chinese exports rising in sectors even when espionage is unsuccessful. This implies that the appropriate counterfactual for China is a rising baseline, meaning the true negative causal effect of espionage on Chinese exports is likely larger than our raw estimate—our detected effects on China are therefore if anything understated.¹⁷ Our conclusion that the US and the rest of the world suffer aggregate losses from espionage is nonetheless well validated by the clean placebo for those exporters.

We note two additional reasons that effects may be large. First, many targeted firms are large, multiproduct firms, and so losses in affected HS6 codes may be larger than the average losses the firm faces. Second, on a ten year horizon, the loss in revenue for targeted firms is similar in magnitude to the decline in exports. Given the size distribution of firms follows a power law and that many sectors are repeatedly targeted, this can reconcile the large estimated effects on exports.

Firm Responses to Espionage

The previous section has established the direct effect of espionage on firm and industry outcomes. Having documented these substantial harms, we turn to a natural follow-up question: how do targeted firms manage their innovation in response to espionage? We study firm responses across four margins: patenting behavior, employment of foreign scientists, participation in international markets, and corporate speech. Restrictions along any of these dimensions could compound the direct economic damages from espionage by reducing the knowledge flows that sustain long-run innovation.

Patent-Based Openness

One way to understand firm openness and knowledge-sharing with the world is through patenting behavior. Patenting inherently involves disclosure of inventions, providing knowledge spillovers to the rest of the world. Patenting data additionally includes information about how the technology was produced (e.g., the location of collaborators) that indicate how internally open the firm is. Thus, we can get a richer understanding of how espionage affects knowledge diffusion by measuring patenting outcomes.

We estimate Equation [eq:twfe] with patent-based outcome measures, with two modifications. First, because we are especially interested in how firms respond to geopolitically sensitive incidents and US-China competition, we restrict our focus to the cases where China is the beneficiary of espionage. Second, we define “treatment” as when espionage against a firm ends, rather than when it begins.¹⁸

A natural concern with adapting Equation [eq:twfe] is that now our treatment is an information shock, not a technology shock—and untargeted firms can respond to this information shock. Put differently; Airbus cannot (directly) lose revenue because Boeing was targeted for espionage, but Airbus can restrict its knowledge sharing in response to Boeing being targeted for espionage. By comparing the response of targeted and untargeted firms, we are implicitly assuming that untargeted firms cannot respond to the information shock provided by espionage. We address this concern by directly testing it. In 8.2, we show that untargeted firms do not change their innovation management in response to espionage against their competitors. Thus, this exercise can recover the effect of espionage on innovation management at targeted firms.

We first estimate the effect of espionage on the propensity for firms to patent. In Figure 18, Panel A, we show that firms targeted by espionage are about 20-30% less likely to patent in the 5-10 years following an espionage incident. This decline in innovation is consistent with the overall declines in firm size, spending on R&D, and number of scientists employed that we document above, and may represent a substantial loss of knowledge given the frontier research that these firms tend to engage in.

However, it is important to distinguish between a reduction in patenting because of scale effects (less revenue and liquidity to fund R&D), and a reduction in patenting because firms are being more secretive about the technologies they are developing. Only the latter contributes to a reduction in knowledge openness. We proxy for this measure of “secrecy” by measuring patents filed per dollar of R&D expenditure—reflecting the fact that if the results of R&D were instead held as trade secrets, that would decrease the patents filed per R&D dollar. In Figure 7, Panel A, we estimate how espionage affects the patent-to-R&D ratio. The results show that if anything, espionage increases the patent-R&D ratio. For firms to have become more secretive, they would have had to become much more efficient in converting R&D spending into real innovation, which is unlikely given the other measures of firm decline. Thus, we do not see firms becoming less likely to share knowledge through patenting after being targeted for espionage.¹⁹

Consistent with this surprising result, we find no evidence of firms restricting their international collaborations across other dimensions. In Figure 7, Panel B, we show that firms are no less likely to patent with inventors living in China, indicating that firms may not adjust their international scientific collaboration practices in response to espionage. The innovation literature also uses patent citations as an indicator of knowledge diffusion. In Figure 7, Panel C, we find that Chinese firms are no less likely to cite patents from the targeted firm in response to espionage. In Figure 18, we examine two other outcomes of interest. Firms are no less likely to patent with inventors living outside the US, and citations to the targeted firm’s patents also do not change in response to espionage. This indicates that targeted firms do not seem to restrict other firms’ awareness of their technologies.

Collectively, these results indicate that targeted firms do not deliberately make their patented technologies less open to China or the rest of the world. This implies that firms do not view patented technologies as providing a large source of espionage risk, relative to the benefits they get from patenting and having a certain set of inventors.

Employment Responses

The key vector for espionage is employees of the targeted firm. Thus, an important margin of response for targeted firms is how they adjust their employment decisions. In particular, we ask: do firms become more discriminatory towards employees who they perceive as “high espionage risk”? In espionage cases involving China, the most likely category that firms would want to restrict is Chinese scientists—especially those doing deep technical work that they could take to Chinese firms. This kind of employment discrimination would be illegal, but is certainly a plausible response by firms.

OPT Hiring

The US immigration system offers work authorization for student F-1 visas, known as OPT (Optional Practical Training). In STEM fields, these may last up to three years (whereas they last only one year outside STEM). As such, the number of OPT employees at a firm is a good proxy for the number of new foreign scientists that it is hiring.

We link firms in our dataset to OPT records provided by , obtained via FOIA request from U.S. Immigration and Customs Enforcement (ICE). We identify Chinese OPT employees by those with either their country of birth or their country of citizenship being China, including Hong Kong. We restrict our focus to individuals with masters/doctoral degrees, in a subject on the list of CIP codes that the Department of Homeland Security classifies as STEM. This is the group of employees that could represent the greatest potential vector for espionage—highly technical talent whose visa status in the US is not certain. As such, if firms respond to espionage by screening out employees with high perceived “espionage risk”, we would expect to see reductions in the hiring of Chinese STEM OPT workers.

In Figure 7, Panel D, we estimate how employment of Chinese OPT workers per million dollars of firm revenue changes following the revelation that espionage has occurred. We normalize by firm revenues to avoid a mechanical effect, whereby firms hire fewer workers overall due to lower revenues and thus also hire fewer Chinese OPT workers. The outcome is standardized to have mean 0 and a standard deviation of 1. We estimate a tight null, able to generally rule out effects larger than 0.02 standard deviations in either direction. In Figure 15, we show that this result is not sensitive to alternative definitions of Chinese OPT workers (including all degrees, all majors, or restricting to mainland China). Together, these results suggest that firms do not respond to espionage by reducing their hiring of Chinese scientists.

LinkedIn Employment

While Chinese OPT workers are plausibly the group most likely to be affected by hiring restrictions following espionage, they are only one group of workers. To gain a broader picture of how firms change their hiring practices following espionage, we use LinkedIn data provided by Revelio Labs. This data captures the stock of all employees at targeted firms, and allows us to examine whether broader employment patterns reflect discriminatory adjustment. Revelio classifies employees into ethnic categories based on their name, allowing us to observe the number of ethnically Asian employees at each firm over time.²⁰ Revelio also classifies job titles into 1500 employment categories, which we manually classify as “scientist” and “non-scientist” jobs to determine whether a job involves R&D work (e.g. “process development engineer”, “research scientist”). This measure is more coarse than the OPT results that make up our primary analysis, but it is more representative, so we see it as a useful complement.

Figure 19, Panel A, shows that while firms see 40% lower employment after espionage incidents (consistent with the overall decline in firm revenue), this decline is uniform for Asian and non-Asian employees. One discordance comes from focusing on scientists—in Panel B, the number of non-Asian scientists falls by 30% five years after an espionage incident, but the number of Asian scientists falls by 50%, with a marginally statistically significant difference (p=0.068).

Unlike our OPT results, this pattern could suggest that firms are screening out ethnically Asian scientists in response to espionage concerns. However, it is also consistent both with discrimination with ethnically Asian scientists having better outside options, that lead them to leave the targeted firm at higher rates—recall that firm revenues and R&D spending fall after espionage, so employees could have good reasons to leave.

To distinguish between these two explanations, we estimate the effect of espionage on diversity and inclusion within targeted companies. Revelio records the text of employee reviews of each firm, from which we construct DEI scores that measure how positively employees speak about the inclusivity of their work environment. Details on how this data is constructed can be found in 7.2. If being targeted for espionage led firms to discriminate against Asian scientists, we might expect to see lower DEI scores based on reviews from scientists, especially compared to non-scientists for whom we do not see this differential selection pattern.

Figure 16 shows that after espionage, there is no decline in the DEI scores of that firm, either by scientists or by non-scientists. Thus, we conclude that the differential decline in the employment of Asian scientists is likely due to those scientists having stronger outside options (and/or greater anticipation of firm difficulties), rather than due to discrimination.²¹ Our LinkedIn results thus line up with the OPT results, in showing that firms do not respond to espionage by changing their willingness to hire Chinese scientists.

Geographic Markets

We use segment data from Compustat that describes the geographical markets in which a firm operates, and their sales in each market.²² Firms may respond to espionage by reducing their global exposure, focusing efforts on certain markets or excluding other markets; they may also expand their coverage in an attempt to diversify.

In Figure 7 Panel E, we find no evidence that firms contract their geographical footprint in response to espionage; the number of markets that firms operate in does not fall after espionage. We also directly examine whether firms change whether they operate in China, the country most associated with espionage in our sample. If firms respond to Chinese espionage by pulling back from China as a market, we would expect a decline in the probability of having a Chinese business segment. In Figure 21, Panel A, we find no evidence that firms exit the Chinese market in response to espionage targeting. One may also think that discrete changes in geographic markets are too large of a response to detect: firms may reallocate operations across geographic markets, while maintaining all preexisting global ties. We examine whether these intensive margin allocative effects occur by testing whether the HHI of sales across geographic segments changes after espionage. In Figure 21, Panel B, we find no systematic evidence of this occurring. This indicates that firms are not strategically responding to espionage incidents by altering their geographic business strategy.

Firms may also respond to espionage across their business operations. We next examine whether firms change their operations across business segments (major product lines or divisions within the company).²³ In Figure 21, Panel C, we track the number of business segments a firm operates in. A firm may strategically narrow its product lines in response to espionage, either to focus resources or to reduce the surface area of potential theft. We find no significant change in business scope by this measure. In Panel D, we test for intensive margin changes by looking at the HHI of sales across business segments. We also find no evidence of changes in their business strategy along this margin.

Earnings Calls

We do not find that firms respond to espionage by adjusting their patenting activity, employment, or geographical coverage. However, the space of possible actions that a firm could take in response to espionage is immense—it would not be possible to directly test for them all. To get as comprehensive of a view as possible, we instead test for changes in the nature of corporate speech, specifically focusing on earnings calls.

We test for changes across the high-dimensional space of corporate speech using the method of , which uses LLM representations of concepts to summarize text data and test for differences in any concept contained in the text. This method allows us to (in principle) pick up on any dimension of a response to espionage represented in public speeches by corporate executives and the analysts questioning them.²⁴ This method is designed to be automated, minimizing researcher degrees of freedom, while allowing for a very large number of hypotheses to be tested with appropriate multiple hypothesis correction. More details on the method can be found in 7.3.

Figure 22 shows the results of this analysis. Of 16,000 tested hypotheses for differences in corporate speech between successfully targeted and unsuccessfully targeted firms, we find that only two differences are significant. Both of these represent stylistic differences in speech and are not economically meaningful ("patterns of formatting changes or markup language" and "the word ’while’ in various contexts"). Thus, earnings call data provides a rich null result, ruling out effects on a wide range of firm responses that would be reflected in earnings call transcripts. This null result is corroborated by our own qualitative reading of earnings calls by these firms; when we read earnings call transcripts in the aftermath of espionage, neither firm representatives nor analysts questioning them tend to discuss the news of espionage and how it affects the firm, even after the espionage has become public knowledge.

Aggregate Firm Openness Index

One possible concern with interpreting our null results is that given our small sample size, a null result could reflect either a true null effect or simply a lack of power in our sample. To address this concern, we aggregate the outcomes tested in the sections above into an index of "firm openness", and test for effects on this index. We define a firm openness index by standardizing each of the five outcomes captures in Panels A-E of Figure 7. We then define the firm openness index as the inverse-covariance-weighted average of these standardized outcomes .

Panel F in Figure 7 shows the results of this analysis. Interpreting the index as a measure of firm openness, we can rule out espionage decreasing firm openness by more than 0.04 SDs in the eight years after espionage ends. This null is much more precise than its constituent outcomes, and provides clearer evidence that our results reflect true non-responsiveness by firms to espionage, rather than limitations of our data.

Discussion

The preceding analysis establishes that across a variety of outcomes—patenting, employment of Chinese scientists, geographical market presence, and corporate speech—firms do not respond to espionage by becoming less open or more restrictive in their innovation management. Given the large economic damages we document, this is a striking and puzzling result.

In standard models of innovation , firms constantly look over their shoulders, monitoring competitors and adjusting strategy in response to threats to their technological lead. What we document is nearly the opposite: when hit by a large negative shock that demonstrably erodes their competitive advantage, firms appear to bury their heads in the sand. This is reinforced by the corporate speech evidence—not only do firms fail to change their behavior, they also fail to even discuss the espionage in earnings calls with investors, even after the cases become public.

Several mechanisms could explain this inertia. First, firms may deliberately minimize public acknowledgment of espionage to limit damage to shareholder confidence; in this view, the absence of any strategic response is itself strategic, aimed at projecting stability. Second, principal-agent problems within firms may prevent adjustment: the managers closest to the affected projects may have strong incentives to downplay the threat rather than flag a failure that could affect their own careers, making it organizationally difficult for firms to mount a coherent response. Third, firms may simply hold an overoptimistic belief in the durability of their technological lead—believing that even if a competitor has stolen their technology, their own accumulated capabilities and ongoing investment will prevent them from being overtaken. Whatever the underlying mechanism, the combination of large economic harm and near-zero behavioral response constitutes a form of corporate inertia that has potentially significant consequences for aggregate innovation and national competitiveness.

Is it possible that the margins we study (patenting, international collaboration, hiring of foreign scientists, international business operations) are simply not important margins for firms to adjust on? It could be that firms see these margins as low-risk vectors for espionage, and instead adjust on margins that we do not observe (e.g. internal access controls). This is unlikely, given that a large literature documents knowledge flows through patenting, scientific collaborations and business operations as being central to innovation.

Another explanation for the non-response could be that espionage does not constitute a genuine information shock. If firms already know that their industry is frequently targeted for espionage, the discovery that they themselves have been targeted need not trigger any response if their security decisions were already optimal.²⁵ Thus, any innovation restrictions that could have negative welfare consequences—e.g. restricted collaborations, employment discrimination—are costs incurred by the mere threat of espionage, even beyond the costs of each espionage incident.

Conclusion

This paper provides systematic evidence on the economic damages from espionage to US firms and industries. We show that economic espionage has substantial negative effects on targeted firms, with revenues and R&D expenditures declining by roughly 40% within five years, with effects persisting for up to a decade. These damages are concentrated in cases where firms claim high economic losses and in industries with lower R&D intensity, suggesting that espionage is most harmful when it targets key technologies that are difficult to replace. Importantly, we find no effects when espionage attempts are unsuccessful, providing strong evidence for a causal interpretation of our results.

These firm-level effects translate into measurable aggregate effects on US industry. American exports in targeted sectors decline by roughly 40% over a decade. Chinese exports in these sectors do not appear to gain through international markets in response to espionage, suggesting that the damages to the US are not offset by observable gains to China through this channel, with global exports in targeted sectors declining by 10%. These results demonstrate that the economic consequences of espionage extend well beyond the individual firms targeted, and are not simply a redistribution of market share among US firms.

Given these substantial economic damages, a natural question is whether firms respond by restricting their knowledge sharing in ways that could further harm innovation and growth. Across a wide range of outcomes, we find no evidence of such responses. Targeted firms do not reduce their patent-to-R&D ratio, do not become less likely to patent with inventors in China or outside the US, and do not see reduced citations to their patents. We also find no evidence of discrimination in employment decisions. While the number of Asian scientists at targeted firms declines differentially compared to non-Asian scientists, this pattern appears to be driven by Asian scientists having better outside options rather than by firm-side discrimination, as evidenced by the lack of decline in diversity and inclusion scores at targeted firms.

These findings have important implications. First, from a policy perspective, our results establish that economic espionage causes substantial economic harm to US firms and industries, validating policy concerns about espionage. The aggregate effects on exports demonstrate that these damages affect broader US competitiveness in targeted sectors, as well as negative spillover effects abroad. Second, our results provide evidence on the effects of competition on innovation. Economic espionage represents an exogenous shock to competitive pressure faced by industry leaders, and our finding that it reduces R&D spending provides support for the Schumpeterian channel whereby increased competition reduces innovation incentives.

Third, despite these substantial damages, we find no evidence that firms respond by restricting knowledge sharing in ways that could create additional harm to innovation and growth. This suggests that concerns about geopolitical tensions leading to widespread restrictions on knowledge diffusion may be overstated, at least in the private sector response to espionage incidents. The fact that firms do not reduce patenting or international collaboration in response to espionage is particularly notable given the magnitude of the damages we document.

Several important questions remain for future research. First, our results raise the question of why firms do not appear to restrict knowledge sharing in response to espionage despite the substantial damages they experience. One possibility is that firms recognize that the benefits of open innovation and international collaboration outweigh the espionage risks. Another possibility is that the knowledge embodied in patents poses relatively little espionage risk compared to other forms of proprietary information, raising broader questions about the role of patents in knowledge disclosure.

Second, while we have compiled the most comprehensive dataset on economic espionage that we are aware of, our dataset represents only the tip of the iceberg when it comes to economic espionage. We cannot make claims about the welfare effects of espionage using the heavily selected sample we observe. Future research could use confidential data from security firms or governments to gain a more representative picture of economic espionage events, and thus be able to make more confident claims about the typical espionage case than we can in our setting.

Third, much knowledge diffusion, both licit and illicit, happens through the movement of workers. Future research could examine the movement of workers between American and Chinese companies to understand the knowledge diffusion embodied in these moves. While our paper is about the harms from economic espionage, it is not an accounting of any costs and benefits associated with the movement of knowledge workers across countries, and such movement can be tremendously beneficial for knowledge production in both countries.

Finally, a caveat is in order for interpreting our results. In examining the economic impacts of espionage, we are aware of the political and racial undertones of discussions about Chinese economic espionage. It is important to highlight that most corporate espionage cases do not involve a foreign country at all, but instead come from domestic competitors . In our dataset and analysis, we focus on foreign economic espionage and Chinese economic espionage in particular, because Chinese economic espionage is an active policy issue of large importance. But there is a fine line between countering Chinese espionage and persecuting ethnically Chinese researchers. As and show, Economic Espionage Act prosecutions are systematically biased against ethnically Chinese defendants. Our findings should not be taken as a justification to lower the standards of evidence for alleging economic espionage in particular cases.

Figures and Tables

Note: This figure presents a timeline of economic espionage incidents against firms in the United States since 1995. Each row contains a separate incident. The shaded bars span the duration in which espionage occurs (the beginning of the bar corresponds to when espionage began, and the ends of the bar correspond to when the proprietary information was last successfully extracted, or when the suspect was caught if they were caught). The announcement date indicates the date in which the espionage incident was first announced to the public. Incidents that are primarily cyber in nature are plotted in red, while non-cyber incidents are plotted in blue.

Note: This figure presents sample means and 95% confidence intervals of the share of industries that are dual-use, split by whether they are targeted or untargeted by espionage. Industries are classified as dual-use or not based on .

Note: This figure plots the rank of victim firm revenue within their industry (SIC code) in red, and the size of their industry in gray. Firms are ordered by rank, and then industry size. The y-axis uses a log-scale. Revenue is calculated as the average value prior to an espionage incident taking place.

This appendix contains additional material, including figures and tables.

Data Cleaning

Scientist Job Classification

Revelio Labs categorizes all LinkedIn job titles into 1500 broad jobs. We manually code jobs as being “scientist” jobs based on their title. Any category whose title contains the word “research” or the word “scientist” is classified as a scientist job (so that, for example, “Senior Scientist” and “Principal Scientist” are captured by the latter rule). Additionally, the following specific categories are also classified as scientists: analytical chemist, automation engineer, CAD designer, CAD engineer, chemical engineer, chemist, clinical research, electrical design engineer, electrical engineering, environmental engineer, environmental scientist, geotechnical engineer, hardware design engineer, hardware engineer, industrial designer, industrial engineer, industrial engineering, innovation, lab, lab tech, lab technician, laboratory, laboratory analyst, laboratory technician, materials engineer, mechanical design engineer, mechanical designer, process developer, process development, process development engineer, process engineer, process engineering, product design, product design engineer, product designer, product developer, product development, product development engineer, product engineer, product engineering, R&D, R&D engineer, R&D project, and scientific.

Employee Reviews on Diversity and Inclusion

RevelioLabs provides individual review data for many of the companies in our sample. These reviews include both an individual’s overall rating of the company, and also the individual’s rating of the company on diversity and inclusion. Additionally, the full text of the employee’s review is also recorded.²⁶ While coverage of the overall ratings and the text of employee reviews range from 2008 to 2025, the diversity and inclusion reviews only begin in 2020. This poses a substantial sample restriction.

To overcome this restriction, we make use of the text of the reviews and natural language processing techniques to impute diversity and inclusion ratings. This allows us to extend coverage of these ratings back to 2008.²⁷ We conduct this imputation in two separate ways: TF-IDF and ridge regression, and a fine-tuned BERT model. Heuristically, the TF-IDF and ridge regression method identifies words that are highly predictive of the ratings and uses a penalized regression on these words to estimate the final rating, while the fine-tuned BERT model involves tuning a language transformer model to our particular text corpus to predict the final rating. We describe each method in more detail.

TF-IDF and Ridge Regression:

We vectorize our training set (data from 2020 onwards) corpus using TF-IDF (term frequency-inverse document frequency). When conducting TF-IDF, we first convert all words to lower case and remove accents. To build the vocabulary of terms, we include both unigrams and bigrams, filter out any terms appearing in over 95% of ratings as stop-words as well as any terms that appear in less than 5 ratings. We also limit the vocabulary to contain a maximum of 300,000 terms. Finally, we use instead of using the raw term frequency, we use 1 + log (term frequency) (sublinear scaling) to smooth the distribution.

With the training set vectorized using TF-IDF, we then run a ridge regression on the diversity and inclusion ratings using λ = 1 (where λ is the constant on the L2 regularization term). This allows us to fit the model on the training data. To complete the imputation for the full dataset, we vectorize the full dataset and apply the fitted ridge regression model to it.

Fine-tuned BERT model:

We begin by splitting the data into 5 folds (with an equal number of companies per fold), splitting the training/test set based on these folds. We next tokenize the text, and use as our baseline model a multilingual BERT (Bidirectional Encoder Representations for Transformers) model fine-tuned for consumer sentiment, nlptown/bert-base-multilingual-uncased-sentiment We then fine-tune the model with a regression head to predict the diversity and inclusion rating. To do so, we use the optimizer AdamW with a learning rate of 2e-5, 3 epochs, and a batch size of 8 for both training and evaluation.²⁸ With the model fine-tuned, we generate predictions for the diversity and inclusion ratings on the full dataset.

On a five point rating scale, we achieve a MAE (mean absolute error) of 0.659 using TF-IDF and ridge regression, and a MAE of 0.728 using the fine-tuned BERT model.

For robustness, we conduct an imputation of the overall ratings using the same methods as the diversity and inclusion ratings. We achieve a MAE of 0.491 using TF-IDF and ridge regression and a MAE of 0.507 using the fine-tuned BERT model. Event study results using overall ratings as the outcome are highly similar regardless of whether we use the actual overall ratings, or the imputed ratings from either method.

Automated Discoveries from Earnings Calls

Do firms communicate in systematically different ways following successful espionage incidents on earnings calls? We use the method in to simultaneously test over 16,000 hypotheses. This method is designed to be automated, minimizing researcher degrees of freedom, while allowing for a very large number of hypotheses to be tested. It also appropriately statistically controls for false discovery rates. Beneath, we summarize the method (including a difference-in-difference extension we develop for our setting), and then discuss technical implementation details.

Method

We begin with the text data of earnings calls from US public companies that are ever targeted for espionage (successfully or unsuccessfully). This text data is passed through a large language model (LLM) equipped with a pretrained sparse autoencoder (SAE). The sparse autoencoder contains approximately 16,000 features, designed to span the underlying concepts that a LLM uses in a human interpretable way. Examples of activations may be substantive (e.g., “references to scientific authors or collaborators, particularly in the context of publications or citations ”, “terms related to emergency medical services and their operations ”, “materials or methods related to chemical analysis and spectrometry”.) or stylistic (e.g., “numerical values related to sizes or dimensions”, “language constructs related to functions and their definitions in code ”). At this stage, the data are a large matrix Y where rows i correspond to a unique earnings call and columns j SAE features, with cells Y_ij a binary indicator of a particular concept being expressed in that earnings call.

constructs a test statistic for differences in means between two groups. We extend it to our differences-in-differences setting, constructing the test statistic as follows: $$T_{f,j} = \frac{1}{\sqrt n} \sum_f \frac{W_f - \pi}{\pi(1-\pi)} \frac{1}{|F|}\sum_{i \in F} [D_{ft(i)} Y_{ij} - (1-D_{ft(i)})Y_{ij}]$$ where f indexes a firm, F the set of calls corresponding to firm f, n the number of rows of Y, W_f an indicator for treatment (1 for successful espionage, 0 for unsuccessful), π the proportion of treated firms, and D_ft(i) = 1 if the call is post-espionage attempt and 0 otherwise. This applies a Horvitz-Thompson transformation after differencing out any within firm fixed characteristics. This also differences out any secular changes in language between pre- and post-treatment periods. An important note is that in this analysis, the untreated group is defined as the firms that were unsuccessfully targeted for espionage—not firms that were never targeted for espionage, as in our main analysis.

All features are then simultaneously tested using the high-dimensional multiple hypothesis testing procedure described in . This is a bootstrap procedure that produces a critical value for each feature while controlling the k-familywise error rate. Thus, from this procedure we identify (1) features that are statistically different across successful and unsuccessfully targeted firms (before and after treatment), (2) point estimates for how much more frequently these features are used, and (3) confidence intervals for each feature.

Implementation

We begin by truncating each document to the first 16000 characters (avg. 3325 tokens).²⁹ We use as our LLM/SAE Gemma Scope 2B.³⁰ Human interpretable descriptions for each of the roughly 16,000 features are obtained from Neuronpedia.³¹

After constructing the test statistics, we obtain critical values and confidence intervals using the bootstrap procedure with k = 5 (for the k−familywise error rate), α = 0.05, and 100,000 bootstrap draws. The critical value is T = 3.909. Across 16,350 features, we reject the null (treatment effect is 0) for 2 features: feature 16223 (‘patterns of formatting changes or markup language elements’, T = 4.190) and feature 4836 (‘the word "while" in various contexts’, T = 4.054).

Additional Analysis

Effects on Targeted Firms

Robustness Checks

We check whether alternative difference-in-difference estimators yield different results from the baseline specification in Equation [eq:twfe]. In Figure 9, we plot dynamic treatment effects under a variety of alternate estimators and show that results are not sensitive to the choice of estimator. This is natural, because recent event study estimators vary primarily in how they compare units treated at different times. However, in our setting, most firms in our data are never treated. As a result, most of the identifying comparisons are between ever-treated and never-treated firms, which the TWFE estimator and more recent event study estimators handle similarly. Therefore, for the rest of the paper we report results only from the TWFE estimator in Equation [eq:twfe].

Secondary Outcomes

While revenues and R&D expenditures are flow quantities, one can also test whether their stock analogues, total assets and intangible assets, are affected by espionage. In Figure 10, Panels A and B, we find that economic espionage has a negative effect on a firm’s total assets and intangible assets. We once again find a lack of pre-trends, and find that their path follows a similar pattern to revenues and R&D, declining by 30% (assets) and 60% (intangible assets) within 4 years of espionage beginning, with effects persisting in the 4 years that follow. However, in Panels C and D, we find noisy null effects on gross/net profit margins, consistent with the interpretation that these firms are declining in scale, but not experiencing changes in mark-ups.

Using operating segment data from Compustat, we can also decompose these changes in firm revenues into differences by exports and domestic sales to see where firms are most affected by espionage. In Figure 12, we show that domestic sales and exports to China both drop following economic espionage incidents. Five years after espionage, domestic sales and exports to China have fallen by 60%; exports to markets other than China are noisier.

Spillovers within Targeted Industries

What happens to other firms in an industry when one of them is targeted for espionage? This could be an identification concern with Equation [eq:twfe]. To address this question, we estimate a triple difference specification using Equation [eq:spillover_ddd], in Figure 11. Overall, we find no evidence of either positive or negative spillovers to other firms within the same SIC 4-digit code.³²

Effects on Stock Returns

To measure effects on firms that might not be captured by the balance sheet data, we examine the stock market response to espionage. To do this, we estimate each firm’s cumulative abnormal returns at various time horizons using Fama-French factors.³³

Figure 13 displays these effects, for two definitions of treatment—when espionage begins (Panel A.1), and when espionage is announced (Panel A.2) We find that the declines in firm valuations also track the declines in firm fundamentals, with a roughly 20 to 30% drop in firm value two years after espionage begins. We interpret the initial drop in a firm’s valuation as the market reacting to a fundamental decline in firm value due to espionage—most of the time, the public does not know that an espionage incident has occurred until after the 2 year window displayed.³⁴ The announcement of an espionage incident is associated with a further 20% penalty to firm valuation in the 2 years afterwards, although this likely combines the fundamental effect of espionage itself alongside any announcement effects. To narrow down the effect of announcement, we examine a firm’s CAR in the days immediately surrounding the announcement in Figure 23, Panel A. Estimates are noisy, but point towards a roughly 4% decline in the week after announcement.

In Figure 13, Panel B, we plot cumulative abnormal returns for firms that are victims of an unsuccessful espionage incident. We find that estimates are very noisy but positive. If those conducting espionage have any inside knowledge (for instance, because they are employees of the victim firm), they are likely to target firms that are soon to outperform. This suggests that, if anything, the effects reported above may be underestimates. However, other interpretations are compatible with the rise in firm value, particularly after an unsuccessful espionage incident is announced: the market may take this as a signal that the firm has valuable technology (worthy of being stolen) or update on a firm’s security capabilities.

Mean Reversion and Industry Dynamics

Figure 3 shows that firms that are targeted for espionage overwhelmingly tend to be leaders in their field—more than half of targeted firms are in the top 3 firms in their industry when they are targeted, often in very large industries. This fact raises concerns that our main specification could be confounded by industry dynamics; if espionage has no effect, but leader firms naturally decline compared to other firms in the same sector, then we might estimate that espionage has negative effects simply because of this mean reversion. To address this concern, in Figure 14, we show that industry leaders do not systematically mean-revert, especially when we focus on firms targeted by espionage. Importantly, this figure is centered on the quarter a firm first becomes an industry leader—often many years before any espionage incident—so it documents pre-espionage revenue stability among leaders rather than post-espionage performance.³⁵

A related concern relates to whether industries targeted for espionage look systematically different from other similar industries, and thus may be selected, or simply have differences driven by idiosyncratic industry dynamics. If so, this could affect the interpretation of both industry level regressions (such as the analysis of trade flows), and the assignment of adjacent industries as never-treated firms. In Table [t:comp_leader_dyn], we test whether industries targeted by espionage exhibit similar dynamics to SIC-3 industries, and find that they are comparable, with no statistically significant differences in the persistence of industry leadership (columns 1-4), size difference between top firms (column 5), or HHI (column 6) between them.

Knowledge Diffusion Spillovers to Other Firms

Targeted firms are not the only firms that could endogenously respond to espionage. For example, Boeing might reduce its willingness to patent with Chinese inventors in response to espionage against Airbus. This is not only an independently important margin of response, but also an identification threat to the previous sections—by taking non-targeted firms as never-treated, we risk underestimating the endogenous response to espionage that could be common across firms.

One way to estimate this aggregate margin of response would be to aggregate all of our outcome measures to the industry level, and estimate the effect of espionage in an industry. Unfortunately, industry-level data is too coarse to have any power, so we cannot run informative regressions at the industry-level. Thus, in this analysis, our never-treated group is firms in nearby SIC 4-digit codes—i.e. firms in the same SIC 3-digit code k(i) but not in the same SIC 4-digit code j(i). Let D_{j(i), t + k} = 1 if firm i is in an SIC j(i) that has been targeted for espionage, and let l(i) denote the SIC-3 code of firm i. Then the estimating equation is

$$\begin{equation} y_{i,t} = \sum_{k = -4}^{10} \gamma_k D_{i,t+k} + \sum_{k = -4}^{10} \delta_k D_{j(i),t+k} + \alpha_i + \alpha_{l(i),t} + \varepsilon_{i,t} \label{eq:spillover_ddd} \end{equation}$$

Here γ_k captures the direct effect of espionage on victim firms, while δ_k captures the spillover response of competitor firms in the same targeted industry. Note that this specification includes SIC-3 by year fixed effects α_l(i), t, in contrast to the SIC-4 by year fixed effects used in Equation [eq:twfe]. This reflects the fact that our never-treated comparison group here consists of firms in nearby SIC-4 codes within the same SIC-3 industry. The competitor spillover effects we report in Figures 20 and 17 come from the δ_k coefficients; victim firm effects are captured by the separate γ_k term.

For the employment outcomes, since Revelio Labs only provides NAICS codes and not SIC codes for firms in the LinkedIn data, we assign treatment at the NAICS 6-digit level, and we use firms in “nearby” NAICS 4-digit codes as the never-treated group. We only keep the first incident targeting an industry in estimation.

In Figure 20, we estimate the spillover impacts of espionage to the employment of other firms in the targeted industry. Panels A and B reveal that firms in industries targeted by espionage also appear to hire slightly fewer employees, in the 5-10% range, among both Asians and non-Asians.³⁶ On the other hand, Panels C and D reveal that the hiring of scientists does not change at untargeted firms, both among Asian and non-Asian scientists—suggesting that that untargeted firms do not necessarily update their hiring practices from the occurrence of espionage in other firms in their industry.

In Figure 17, we estimate the spillover impacts of espionage to patenting by other firms in the industry. Panel A shows that firms in industries targeted by espionage are about 5% more likely to file a patent, consistent with the small increases in the hiring of scientists that we observe. However, we find no effects on the likelihood of patents Taken together with the results on employment, these results suggest that firms in industries targeted by espionage may try to increase their innovation activities—potentially to develop a lead over the stolen technology, or to outcompete the victim firm—but that they do not necessarily become less open to foreign collaboration and innovation.

One explanation for this non-response even at the industry level is that firms’ decisions about optimal openness depend on their perception of espionage risk. As such, even if espionage makes them less open to knowledge flows, they might not respond to individual events of espionage. Indeed, given that many sectors are targeted repeatedly (Table [t:targeted_industries]) and thus presumably face a high background espionage risk, this is the natural interpretation to our null findings on the endogenous response of firms. However, our estimation only includes the first time an industry is targeted for espionage. Thus, as long as the first espionage incident in a sector updates firms’ beliefs about espionage risk, our strategy should capture the responses of firms to espionage. The fact that we find no effects on other firms even from the first espionage incident in a sector implies that this is not simply an artifact of firms having generally well-calibrated beliefs about espionage.³⁷

Additional Tables and Figures

Note: This figure plots the distribution of T-statistics (as defined in 7.3) of difference-in-difference tests between successfully and unsucessfully targeted for espionage firms across 16,350 features of earnings calls. The critical value (T = 3.909) is displayed as a red dashed line. The two statistically significant features are ‘patterns of formatting changes or markup language’ (T = 4.191) and ‘the word "while" in various contexts’ (T = 4.054).