Measuring Judicial Bias with Artificial Intelligence

The paper presented in this ZEW Research Seminar studies judicial bias in intellectual property (IP) litigation and its implications for innovation incentives. Exploiting a comprehensive dataset of Chinese IP court decisions from 2014–2020, the authors document a striking puzzle: despite widespread concerns about local protectionism, non-local plaintiffs often exhibit higher observed win-rates than local plaintiffs. To explain this puzzle, they develop a theoretical framework in which local governments’ fiscal incentives generate two intertwined effects. The first is a ''distortion effect''—comprising local protection bias, which favors local firms over non-local rivals, and monopoly bias, which favors local plaintiffs when both parties are local. The second is the ''picket fence effect'', whereby anticipated judicial bias endogenously reshapes the composition of cases that enter litigation. As a result, observed win-rates conflate judicial distortions with selection effects, making judicial bias fundamentally challenging to identify from raw outcomes. To address this challenge, the authors train an LLM–based ''AI court'' on cases in which both litigants are non-local, generating counterfactual fair win-rates for all other disputes. Comparing observed and predicted win-rates reveals significant judicial bias. Leveraging a 2019 reform that centralized appellate jurisdiction over key IP cases at the Supreme Court, we show that stronger central supervision substantially improves judicial accuracy and curtails bias. These institutional improvements translate into sizable increases in firms’ innovation activity. The findings highlight the central role of judicial impartiality in underpinning innovation incentives and demonstrate the value of AI-based counterfactual tools for evaluating institutional distortions.