Corporate AI Safety Governance and Firm Productivity
Description
This study uses A-share listed firms in China from 2019 to 2025 as the initial sample. The starting year corresponds to the establishment of China’s National New Generation Artificial Intelligence Governance Expert Committee and the issuance of the Governance Principles for a New Generation of Artificial Intelligence, which formally introduced AI safety governance as a core topic in national policy and marked the beginning of China’s systematic governance framework. We apply three sample filters, excluding (1) firms designated as Special Treatment during the sample period, (2) firms in the financial industry , and (3) observations with missing values for relevant variables. All continuous variables are winsorized at the 1 percent level to mitigate the influence of outliers. The final sample consists of 18,509 firm-year observations covering 4,375 firms. Patent text data are obtained from the IPRDB Intellectual Property Database. Annual report texts and corporate social responsibility report texts are collected from Cninfo . Official news text data and financial data are obtained from the China Stock Market and Accounting Research (CSMAR) database. For the AI safety variable, two features of the scoring rules warrant note. First, some overlap across risk-point dictionaries is intentional, because a single technical method can address more than one type of AI risk. For example, generative adversarial networks (GAN) can be used to generate minority-group samples for data debiasing (Item 2) and to detect or process inappropriate training data (Item 8). When a firm’s disclosures meet the criteria of two items through such a method, both items are credited, because the index measures the breadth of risk coverage and a technology addressing two risk points represents broader coverage than one addressing a single point; the binary structure caps the contribution of any item at one point. For polysemous terms, disclosure text does not always identify the function a method serves, and multi-item credit may occasionally reflect a single underlying activity. The manual semantic screening of the lexicon, the sentence-level co-occurrence requirements, and the subcategory-breadth conditions limit this possibility, and remaining misclassification introduces measurement noise rather than systematic scoring advantages. Second, the term subcategories within an individual rule are groupings of conceptually distinct evidence used to assess breadth; covering an additional subcategory does not generate an additional point, and only the binary item-level criterion is scored. The detailed Chinese-language scoring rules, together with the complete seed-term list and lexicon, are provided in the online supplementary materials.
Files
Institutions
- Shanghai UniversityShanghai, Shanghai