This paper provides descriptive evidence on how stock market participants use Generative Artificial Intelligence (GenAI) to process investment‐related information. Using a data set of 1.7 million stock‐related queries from one of China's largest GenAI platforms during the first half of 2024, we document that user queries address a wide range of topics and tasks and vary systematically with usage intensity and financial sophistication. Query activity increases around corporate disclosure events, but these increases largely track contemporaneous media coverage. We also find evidence consistent with a substitution between the informativeness of voluntary managerial disclosures and investors' reliance on GenAI. Continued platform engagement more likely follows answers that are concise and contain directionally accurate trading signals. Over time, users' subsequent queries increasingly reflect the specificity and financial terminology present in earlier GenAI answers. At the market level, GenAI usage is associated with higher measures of informed trading and lower liquidity, while aggregated sentiment in GenAI‐generated answers correlates with same‐day abnormal returns, particularly when user feedback is positive. Overall, our findings offer insights into early‐stage GenAI adoption by retail investors and inform discussions on how GenAI shapes information processing in financial markets

I use the setting of generative AI (GenAI) to examine how processing tool interactivity affects investors’ self‐assessments of ability and willingness to invest. Although GenAI can help investors process financial information, I theorize that the interactive nature of GenAI blurs the boundaries between investors’ own abilities and those of GenAI, prompting investors to discount their reliance on GenAI and misattribute its abilities to themselves. I rely on the advantages of a laboratory setting to disentangle the interactive element of GenAI from the mere presence of GenAI assistance. Across three experiments, I find that the interactivity underpinning GenAI heightens investors’ self‐assessments of their own abilities and increases their willingness to invest, despite this interactivity not improving, and in fact hindering, their actual processing of information provided by GenAI. My study thus highlights one potential cost of using GenAI and other highly interactive processing tools.

In this paper, we develop novel measures of managerial labor market classification and competition by constructing networks of compensation benchmarking peers disclosed in proxy statements. These networks represent firms’ relative positions within the managerial labor market. Our classifications strongly predict executive moves across firms, outperforming a comprehensive set of predictors in the literature. Subsequent tests further demonstrate the strength of our methodology in capturing the multidimensional and dynamic features of the managerial labor market. We also validate our competition measures by showing that they are associated with retention tools, such as higher equity pay and longer pay duration. Finally, we apply our measures to test two theoretical predictions. First, we find that labor market competition could explain controversial pay practices. Second, we demonstrate that the labor market provides managers with tournament incentives to deliver superior future performance

This paper explores the influence of two fixed payment arrangements—time‐based and output‐based wages—on worker behavior and performance in a multidimensional task setting. We examine how these wages affect the time workers spend on individual units of a task and their work quality. We contend that fixed compensation schemes can implicitly communicate standards of acceptable work. Our empirical evidence from MTurk experiments and a laboratory experiment indicates that workers on output‐based wages deliver higher quality and spend more time on individual units than their time‐based counterparts. These findings are consistent with output‐based wages, implying a standard of acceptable quality—without a conflicting standard of speed—to which workers respond. Our results emphasize the power of implicit cues from fixed compensation schemes and offer insights for employers, suggesting the choice between output‐ and time‐based wages should be informed by whether quality or turnaround time is valued more

We investigate the monitoring quality of accountants with ties to the Mafia in their role as auditors for “clean” firms—those with no known ties to organized crime. Using a proprietary government database, we identify Italian firms with alleged ties to the Mafia through their executives, directors, or shareholders. We define “suspect accountants” as those who serve as auditors for these Mafia‐connected firms, acknowledging their potential associations with criminal entities. We predict and find evidence that “clean” clients (treatment group) monitored by suspect accountants are more likely to engage in earnings management practices that reduce taxable income, compared with a control sample of “clean” firms monitored by accountants with no known Mafia ties (control group). Our findings suggest that accountants with ties to the Mafia act as low‐quality monitors in the “clean” economy.

This paper hypothesizes that information flows from target firms to large shareholders during activist campaigns and that these flows have governance consequences. Focusing on actively managed mutual fund families, we find that informed trading by large‐holding fund families increases during activist campaigns relative to smaller‐holding fund families invested in the same firms. The effect is stronger for firms that attend more invitation‐only investor events, face greater threats from activist campaigns, and are harder to value. Consistent with information flowing from management to large‐holding fund families, the effect strengthens when Regulation Fair Disclosure enforcement is lax and when the information is favorable to the firm. Furthermore, the increased information advantage is associated with more management‐friendly voting behavior by these investors and a higher likelihood of target firms winning activist campaigns and retaining board seats. Overall, our findings are consistent with a potential quid pro quo in which investors’ access to information from management is associated with more pro‐management behavior

We study the economic consequences of anti‐loss trafficking rules, which disallow the use of loss carryforwards as a tax shield after a substantial ownership change. We use staggered changes to these rules in the EU27 Member States, Norway, and the United Kingdom from 1998 to 2019 and find that limiting the transfer of tax losses is related to the number of mergers and acquisitions (M&A) declining by 18%, driven by loss‐making targets. Turning to broader industry dynamics, we find decreases in survival rates of young companies after tighter regulations. Loosening of regulation is associated with increased firm survival. Tightening (loosening) anti‐loss trafficking rules is related to decreased (increased) industry productivity, especially in R&D‐intensive industries that are more prone to loss‐making. Finally, tighter anti‐loss trafficking rules are associated with lower deal synergies and risk‐taking. All effects concentrate in strict regimes.

Recent work within accounting and finance has highlighted that modern AI systems exhibit superhuman performance on a variety of foundational activities within these fields. However, the literature often does not provide economic rationale for why AI models seem to outperform, largely because these models are a black box. Through a series of experiments, I set out to open the black box and provide direct evidence on how and why AI models appear to perform so well on accounting and finance‐related tasks. I show that much of the superior performance of AI models can be attributed to artifacts of the modeling itself, rather than to mechanisms grounded in economics. Focusing on two key components of AI models, which may bias inferences in papers that rely on them, I first show that Large Language Model (LLMs) exhibit extremely poor numerical reasoning and thus application in these settings should proceed with caution. Second, I highlight that commercial LLMs suffer from significant look‐ahead bias, which may explain a large portion of their predictive ability in various settings

Most loan contracts in the United States contain a provision for lender voting rules. We study the optimal voting rule that allows lenders to waive a covenant violation. When lenders have heterogeneous preferences, lenient voting rules increase the probability of waivers that allow inefficient investments. Stringent voting rules tend to allocate the marginal vote to lenders who deny waivers after false alarms so that they can renegotiate the loan to extract value from the firm, which incurs deadweight costs. In equilibrium, the optimal voting rule balances these two forces to improve contracting efficiency. We derive and empirically test comparative statics on how the optimal voting rule varies with lenders’ preferences and the borrower's accounting properties. Our model offers a rationale for the prevalent use of voting rule clauses in syndicated loan contracts.

We examine the usefulness of machine learning approaches for measuring vocal tone in corporate disclosures. We document a substantial mismatch between the widely adopted actor‐based training data underlying these approaches and speech in corporate disclosures. We find that existing models achieve near‐perfect vocal tone classification within their training domain. However, when tested on actual executive speech during conference calls, their performance declines to chance levels. We thus introduce FinVoc2Vec, a deep learning model that adapts to audio recordings of conference calls and classifies the vocal tone of executive speech significantly more accurately than chance. FinVoc2Vec estimates are associated with future firm performance and can be used to construct profitable stock portfolios. Throughout our analyses, estimates from previous vocal tone models are largely unrelated to firm performance. Our findings emphasize the importance of a domain‐specific approach to voice analysis in accounting and finance.