BETTER AI, WORSE DISCLOSURES? THE UNINTENDED CONSEQUENCES OF NLP ON FINANCIAL REPORTING

This paper examines whether the emergence of Large Language Models ultimately improved or worsened communication between investors and firms. I develop a disclosure model in which investors rely on algorithms to process reports and managers can tailor their language to misdirect the algorithm. The model predicts a non-monotonic relationship between accuracy and informativeness: at the high accuracy levels achieved by LLMs, strategic tailoring dominates processing gains and informativeness declines. Using U.S. earnings conference calls, I show that following ChatGPT-3.5's release in November 2022, managerial tailoring increased while the informativeness of textual reports declined. The decline concentrates in prepared speeches, where optimization is feasible, not in spontaneous Q\&A. Validating the theoretical mechanism, I show that the earlier, less accurate Loughran-McDonald dictionary generated opposite effects: tailoring increased but informativeness rose. The findings demonstrate that algorithmic advances create a fundamental trade-off: improved processing can degrade disclosure quality when higher accuracy enables reliable gaming.