Intellectual Property Judicial Reform, Litigation Bias, and Innovation: An "AI-Court" Approach and Evidence from China

Abstract

The fairness of intellectual property (IP) litigation profoundly influences firms’ incentives to innovate. Due to tournament-style promotion competition among local government officials in China, local courts may exhibit at least two biases in IP dispute rulings: first, local protection bias: the courts may favor local litigants when local and non-local firms are involved in IP disputes; second, monopoly bias: the courts may favor local plaintiffs in IP disputes where both parties are local because tax revenues from the monopolistic profits can exceed that from the combined profits of a duopoly. In our analysis, we define “local firms” as firms that are headquartered in the same province as the court, a notion that is consistent with how the judges in the local courts are appointed during our analysis period of 2014 and after. Testing for these biases is further complicated by a “picket fence” (i.e., selection effect): when plaintiffs anticipate discrimination by the court, they may simply choose not to file a lawsuit if the likelihood of winning is low; as a result, only lawsuits with a relatively high probability of success would appear in the observational data. 

Indeed, when analyzing the data extracted from the texts of initial judgments of cases filed in Chinese courts from 2014 to 2020, as published on the China Judgments Online (CJOL) platform, we find, somewhat puzzlingly, that the plaintiffs have a lower win rate in “local plaintiff vs. non-local defendant” disputes than in “non-local plaintiff vs. local defendant” disputes; similarly, the plaintiffs have a lower win rate in “local plaintiff vs. local defendant” disputes than in“non-local plaintiff vs. non-local defendant” disputes. These raw comparisons are inconsistent with the local protection and monopoly biases, suggesting the importance of accounting for the “picket fence” effect. 

This paper proposes a novel approach that leverages large language models (LLMs) to train a fair “AI Court” on the sample of “non-local plaintiff vs. non-local defendant” cases, for which theory predicts that the courts should be unbiased in their judgment. The predictions of the trained “AI Court” provide the counterfactual fair win rate for the plaintiffs that are free of the aforementioned local protection and monopoly biases, in all the other cases with at least one side being a local firm. By comparing the actual win rates in these cases with the corresponding fair win rates predicted by the “AI Court,” we confirm that the local protection bias and the monopoly bias distort the win rates in favor of local plaintiffs by approximately 1.9% and 3.1%, respectively. We also find evidence of a significant “picket fence effect”, which allows us to reconcile the existence of the local protection and monopoly biases and the apparent counterintuitive raw comparisons we previously mentioned. 

We further study the effect of a 2019 judicial reform where appeals against the initial judgments rendered by the local courts on IP disputes involving “utility models” and “invention” patents were moved from the Provincial High Court to the Supreme Court. We argue that this reform will lead to a reduction in the courts’ local bias and improve the quality of the court’s initial judgment. Using Difference-in-Difference (DID) method and leveraging the trained “AI Court”, we show that this judicial reform reduced the error rate of the initial judgment by the local courts by somewhere between 22.60 and 27.37 percent; moreover, the reform lowered the litigation bias caused by the local protection and monopoly biases by 87.18 percent. 

Finally, we show that the improved quality of IP court decisions as well as the reduced uncertainty in trial outcomes significantly enhances innovation incentives, contributing to about 13.4% and 9.7% increases, respectively, in invention patent filings and utility model patent filings.