PUBLISHED PAPERS

3) Deadly Disease, Fear, and Ostracism in the Credit Market (with Sumit Agarwal, Pengsheng Lin, Ruichang Lu, and Xiaojun Zhang) (2026) Managment Science.

ABSTRACT:  Utilizing the 2014 Ebola outbreak in the United States, this paper examines the impact of fear on credit distortions in the mortgage market. Even though the actual risk of Ebola infection in the United States is extremely low, the prevailing Ebola fear leads to increased rejections for Black mortgage applicants who are perceived to be at an elevated risk of exposure to the virus. Specifically, Black applicants face an additional 2.8 percentage point increase in their denial rate during the Ebola outbreak. We further employ the geographical distribution of the four diagnosed Ebola cases in the United States and algorithm-based lending as a control group to strengthen our analysis of the causal effect of fear on credit distortions in mortgage approval rates. Additional analyses reveal that approved Black applicants tend to have higher FICO scores and lower default probability, and Black borrowers are also less likely to be granted modifications upon delinquency. We also demonstrate a decline in the denial rate of Black applicants immediately after the outbreak, suggesting that loan officers consciously counteract their biases as Ebola fear subsides. Finally, we document that a learning effect translates into a relatively lower denial rate for Asian applicants during the COVID-19 pandemic. 

 2) The Impact of Generative AI on Information Processing: Evidence from the Ban of ChatGPT in Italy (with Jeremy Bertomeu, Yibin Liu, and Zhenghui Ni). 2025 Journal of Accounting and Economics.

ABSTRACT: On March 31, 2023, the Italian Data Protection Authority deemed ChatGPT in violation of privacy laws and banned the service in Italy, providing a natural experiment to evaluate an unexpected increase in information processing costs on capital markets. We employ metrics for AI-generated text detection to show that the ban coincides with decreased utilization by domestic financial analysts and fewer earnings forecasts issued relative to foreign analysts covering the same firm. These adverse effects are more pronounced for analysts whose pre-ban reports were more consistent with AI use or analysts with a technical background. The ban also leads to a reduction in forecast accuracy and a greater reliance on industry-specific information. Further, information efficiency decreases, investors react more strongly to information from earnings announcements, and bid-ask spreads widen.

1) Does Information Processing Cost Affect Firm Specific Information Acquisition -Evidence from XBRL Adoption (with Dong Yi, Oliver Zhen Li, and Ni Chenkai). 2016 Journal of Financial and Quantitative Analysis      

 Best paper award at the 2013 IEFA & CSBF Joint Conference

ABSTRACT: We examine how information-processing cost affects investors’ acquisition of firm-specific information using a natural experiment resulting from a recent mandate requiring U.S. firms to adopt eXtensible Business Reporting Language (XBRL) when submitting filings to the U.S. Securities and Exchange Commission (SEC). XBRL filings make financial data standardized, tagged, and machine readable. We find that XBRL adoption reduces firms’ stock return synchronicity. The reduction in synchronicity mainly applies to filings under the mandatory program as opposed to the voluntary program. Furthermore, such an effect is more pronounced for opaque and complex firms. Finally, we find that XBRL adoption also reduces price delay.

 WORKING PAPERS

4）Automation or Augmentation?  Common Knowledge Frictions and Auditor Pipeline Risk (with Rui Shi and Gaoqing Zhang) 2026 

ABSTRACT: The audit labor market presents a puzzling disequilibrium. As AI increases concerns about automation in auditing, the supply of new accounting talent is contracting: enrollments are falling, entry-level wages remain sluggish, whereas audit firms continue to post more audit positions, leading to an excessive demand. We develop a model in which this pattern arises from higher-order uncertainty about AI-driven technological change. Prospective auditors respond not only to their own beliefs about whether AI substitutes for or complements audit labor, but also to what they believe others believe. When common knowledge about automation dominates common knowledge about augmentation, equilibrium labor supply falls below the full-information benchmark. This coordination failure generates persistent excess demand and reconciles declining enrollments, rising junior-auditor demand, and limited wage adjustment. We further test this model by exploiting AI faculty departures from universities as shocks to the local production of common knowledge about AI’s automation vis-a-vis augmentation value. Using a stacked difference-in-differences design, we find that these departures reduce entry into the audit profession, increase junior audit postings, generate little wage response, and worsen long-run audit quality. The evidence suggests that universities do more than produce skills: they coordinate expectations about the future value of those skills and thereby sustain the pipeline of AI-complementary human capital.

3) AI Information Processing, Misinformation, and Voluntary Disclosure: Theory and Evidence (with Jeremy Bertomeu, Yibin Liu, and Zhenghui Ni) 2026 

ABSTRACT: This paper investigates the impact of generative AI on firms’ voluntary disclosure choices. Our theoretical model highlights a trade-off between AI’s improved ability to process disclosed information and its potential for misinformation, modeled as a random “hallucination” unrelated to the firms’ fundamentals. We predict that increased AI processing leads to more strategic non-disclosure due to two related economic forces. First, hallucinations provide additional camouflage after strategic non-disclosure. Second, because users consider the risk of misinformation, they discount observed marginal disclosures, further reducing the benefit of disclosure. To test our predictions, we leverage OpenAI’s launch of ChatGPT in November 2022 as a shock to AI processing. Consistent with the theory, firms with more AI processing reduce their voluntary disclosures. Further, the introduction of ChatGPT reduces information processing failures, which manifests in increased information processing speed. Combining the crowding-out effect on information supply and the positive impact on information processing speed, we do not find evidence of a net increase in information quality. 

2) The Economic Consequences of Disrupted Generative AI Adoption  (with Jeremy Bertomeu, Yibin Liu, and Zhenghui Ni) 2026

ABSTRACT: Generative artificial intelligence (AI) promises major productivity gains, yet its development is increasingly shaped by policy interventions. We study Italy’s 2023 ChatGPT ban as a natural experiment to assess the economic consequences of disrupted AI adoption. Firms with higher exposure to generative AI experience a 6 percent market value decline during the ban relative to less exposed firms, with stronger effects among smaller and younger firms. These firms face persistently negative media and analyst sentiment, reduced full-time hiring—especially in AI-complementary occupations—and a slowdown in patent filings. We calibrate a Schumpeterian model to match these patterns and estimate that an unimpaired adoption of generative AI increases long-term GDP growth by 0.5 percent. The model reveals redistributional effects on firm value consistent with the empirical results, underscoring the broader economic consequences of disrupted AI adoption for productivity growth and capital markets. 

1) Careful What You Say to AI Ears: A Field Experiment (with zhenghui Ni, Yichen Su, and Nianhang Xu) 2026

ABSTRACT: We run a large-scale randomized controlled trial (RCT) with 2,722 Chinese listed firms to examine how perceived AI bots shape managers’ disclosures and market reactions. In online earnings conferences (OECs), we randomize the questioner’s identity as either human or AI while holding information content fixed. Managers are 13% less likely to answer AI-styled questions, with stronger effects when an image signals AI identity. Linking manager and firm traits to responses shows smaller declines at state-owned enterprises and larger declines for managers who are older, or have a computer science background. Conditional on responding, managers provide longer answers with more neutral narratives to AI-styled questions. Using high-frequency data, we find that investors withhold trading during AI-question windows. Overall, our results are consistent with AI playing a disciplinary role on voluntary disclosure.