Voluntary Disclosure, Misinformation, and AI Information Processing: Theory and Evidence

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 nondisclosure due to two related economic forces. First, AI hallucinations provide additional camouflage after strategic non-disclosure. Second, because users consider the risk of misinformation, they discount observed marginal disclosures, further reducing the payoff to voluntary 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, we find that firms with more AI processing reduce their voluntary disclosures. Furthermore, we estimate Smith's (2024) structural model of earnings announcements and find that the introduction of ChatGPT reduces the information processing failures, manifested in increased information processing speed. Combining the crowding-out effect on information supply by firms and the positive effect on information processing speed, we do not find evidence of a significant net increase in information quality.

Keywords: AI hallucination, Information processing, Information crowding out, ChatGPT