Andrew Kao

Recent scholarship has suggested that artificial intelligence technology and autocratic regimes may be mutually reinforcing. We test for such a mutually reinforcing relationship in the context of facial recognition AI in China. To do so, we gather comprehensive data on AI firms and government procurement contracts, as well as on social unrest across China during the last decade. We first show that autocrats benefit from AI: local unrest leads to greater government procurement of facial recognition AI as a new technology of political control, and increased AI procurement indeed suppresses subsequent unrest. We then show that AI innovation benefits from autocrats’ suppression of unrest: the contracted AI firms innovate more both for the government and commercial markets, and are more likely to export their products; and non-contracted AI firms do not experience detectable negative spillovers. Taken together, these results suggest the possibility of sustained AI innovation under the Chinese regime: AI innovation entrenches the regime, and the regime’s investment in AI for political control stimulates further frontier innovation.

Citizens have long taken to the streets to demand change, expressing political views that may otherwise be suppressed. Protests have produced change at local, national, and international scales, including spectacular moments of political and social transformation. We document five new empirical patterns describing 1.2 million protest events across 218 countries between 1980 and 2020. First, autocracies and weak democracies experienced a trend break in protests during the Arab Spring. Second, protest movements also rose in importance following the Arab Spring. Third, protest movements geographically diffuse over time, spiking to their peak, before falling off. Fourth, a country’s year-to-year economic performance is not strongly correlated with protests; individual values are predictive of protest participation. Fifth, the US, China, and Russia are the most over-represented countries by their share of academic studies. We discuss each pattern’s connections to the existing literature and anticipate paths for future work.

We collect comprehensive data on global trade in surveillance AI technology, and document three facts about its diffusion. First, China has a comparative advantage in this technology, exporting substantially more surveillance AI than other countries, particularly compared to other frontier technologies. Second, autocracies and weak democracies are more likely to import surveillance AI, and more likely to do so from China. Third, autocracies and weak democracies are especially likely to import China’s surveillance AI following domestic unrest. Such imports coincide with broader declines in domestic institutional quality, suggesting that China may be exporting its surveillance state via trade in AI.

Members of Congress do not write their own bills---their staffers do. We estimate the productivity of personal staffers in the US House of Representatives. To do so, we extend the mover design to team settings where an individual's output is unobserved. We find that staffers account for at least 40% of the gap in the number of bills written between high- and low-performing offices. In our setting, conventional estimation methods are either infeasible or misspecified and deliver estimates less than half as large. We analyze three factors related to the optimal composition of teams, finding that managers are less productive than other types of staffers in teams, diverse offices are more productive, and that staffers are more effective when paired with effective Representatives. Finally, we estimate staffer ideology and find that staffers help moderate ideological extremism within parties, while Representatives drive the majority of polarization.

In this paper, we evaluate the capability of large language models to conduct personalized phishing attacks and compare their performance with human experts and AI models from last year. We include four email groups with a combined total of 101 participants: A control group of arbitrary phishing emails, which received a click-through rate (recipient pressed a link in the email) of 12%, emails generated by human experts (54% click-through), fully AI-automated emails 54% (click-through), and AI emails utilizing a human-in-the-loop (56% click-through). Thus, the AI-automated attacks performed on par with human experts and 350% better than the control group. The results are a significant improvement from similar studies conducted last year, highlighting the increased deceptive capabilities of AI models. Our AI-automated emails were sent using a custom-built tool that automates the entire spear phishing process, including information gathering and creating personalized vulnerability profiles for each target. The AI-gathered information was accurate and useful in 88% of cases and only produced inaccurate profiles for 4% of the participants. We also use language models to detect the intention of emails. Claude 3.5 Sonnet scored well above 90% with low false-positive rates and detected several seemingly benign emails that passed human detection. Lastly, we analyze the economics of phishing, highlighting how AI enables attackers to target more individuals at lower cost and increase profitability by up to 50 times for larger audiences.

We provide systematic evidence on the economic damages from espionage to US firms and industries. Compiling a comprehensive dataset of publicly disclosed espionage incidents from 1995-2024, we establish that espionage has substantial negative effects on targeted firms. In an event-study design, revenues and R&D expenditures at targeted firms decline by roughly 40% within five years, with effects persisting for up to a decade. These effects do not appear for firms unsuccessfully targeted for espionage, supporting a causal interpretation. These firm-level damages translate into measurable aggregate effects on US industry: exports in targeted sectors decline by 60% over a decade. Given these substantial damages, we investigate whether firms restrict knowledge sharing in response to espionage. Across a wide range of outcomes, we find no evidence of such restrictions. Firms do not reduce their patenting with foreign inventors, and do not discriminate in employment based on perceived espionage risk. Overall, espionage has clear economic harms to targeted firms and US industry, but firms are puzzlingly unresponsive in how they manage innovation.