The Future of Accounting: AI-driven Tone Manipulation in SEC Filings

Abstract

This paper explores the potential impact of large language model–generated rewritings of 10-K and 10-Q filings. By using LLMs to produce more positively toned versions of these filings, the study compares the sentiment of original versus LLM-rewritten texts using multiple sentiment-analysis techniques, including dictionary-based methods, machine learning models, and LLM-based scores. The results show a notable increase in positive sentiment in the LLM-rewritten filings. Furthermore, among the methods tested, sentiment scores derived from LLMs demonstrate the strongest correlation with stock price movements. A combination of these findings indicates that AI-driven text generation can enhance market perception. These findings carry significant implications for investors and regulators, underscoring the need for careful consideration of AI’s role in financial disclosures.

Keywords: Sentiment Analysis, SEC Filings, Natural Language Processing (NLP), Financial Disclosures, Generative AI, Market Efficiency

JEL Classification: C45, C80, C83, D83, G14, M41