研究会常务理事孙运传最新工作论文介绍
以下文章来源于金融大数据研究中心IFIND,作者孙运传徐荧 曾笑萍

随着ChatGPT等生成式人工智能工具的普及,越来越多的上市公司开始用其撰写年报中管理层讨论与分析(MD&A)部分(Blankespoor et al., 2025)。但北京师范大学金融大数据研究中心孙运传团队的最新研究发现:AI写的MD&A看似流畅规范,却可能暗藏隐患。不仅会降低信息披露质量、推高信息风险,还会直接影响短期股价表现。
这篇基于2021-2025年中国A股4,730家上市公司、20,486条观测数据的研究,揭开了AI在财务披露中的信息风险效应,对投资者、企业和监管机构都极具参考价值。
Abstract: The rise of generative artificial intelligence (GAI) has fundamentally changed the way how firms finish their financial reporting tasks. Using a sample of Chinese A-share listed firms from 2021 to 2025, this paper employs event study to explore whether AI-generated texts in MD&A disclosures shapes information risk and investor responses. We measure the likelihood of AI-generated content (LAIGen) in MD&A narratives via text perplexity based on large language models. It isfound that lower LAIGen in MD&A is positively associated with short-window stock returns around annual report releases. This effect is more pronounced for firms with heightened uncertainty, such as negative unexpected earnings, lower proportions of quantitative disclosure, lower textual burstiness, or weaker external monitoring. Further analyses suggest that AI-generated narratives reduce the firm-specific informational content of disclosures, increase reliance on generic industry- and market-level language, and widen bid-ask spreads, thereby elevating information asymmetry. We address endogeneity concerns using an instrumental variable approach and exploit the exogenous shock of the public release of ChatGPT in a difference-in-differences design, which confirms that reduced AI-adoption costs lead to more widely use of AI-assisted disclosure, particularly among non-AI-industry firms. Our findings highlight the potential unintended consequences of generative AI in firm reporting: while AI may enhance linguistic polish, its use could impair disclosure informativeness, elevate information risk, and weaken market efficiency. The results carry implications for regulators, investors, and firms navigating the evolving landscape of AI-integrated financial communication.
摘要:生成式人工智能(GAI)的兴起,从根本上改变了企业完成财务报告工作的方式。本文选取2021到2025年中国A股上市公司作为研究样本,探究管理层讨论与分析(MD&A)披露中的人工智能生成文本是否会影响信息风险与股票市场反应。本文以文本困惑度为指标,衡量MD&A中人工智能生成内容的可能性(LAIGen)。研究发现,MD&A的人工智能生成可能性越低,年报披露窗口期的短期股票回报率越高。这一效应在企业存在负向未预期盈余、量化披露占比更低、文本突发性更弱或外部监督力度更弱的情况下更为显著。进一步分析表明,人工智能生成的文本会降低公司特质信息含量,增加对通用性行业及市场层面表述的依赖,扩大买卖价差,进而加剧信息不对称程度。为解决内生性问题,本文采用工具变量法,并利用ChatGPT公开发布这一外生冲击构建双重差分模型。研究证实,人工智能应用成本的下降会促使非人工智能行业的企业更广泛地采用人工智能辅助披露。本文的研究结论揭示了生成式人工智能在企业信息披露领域可能存在的非预期后果:尽管人工智能技术能够优化文本的语言表达,但使用人工智能生成披露内容会削弱信息披露的有效性,提高信息风险,降低市场效率。研究结果可为监管机构、投资者及企业提供参考。
原文链接(SSRN):https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5956736