Talk or Deployment? AI Disclosure and the Limits of Text-Based Adoption Measures

Abstract

The share of S&P 500 firms mentioning artificial intelligence in their annual report rose from 38.6% in fiscal 2018 to 96.7% in fiscal 2025. We ask whether this increase measures adoption or disclosure. Using 9,177 paragraphs extracted from Form 10-K filings, we classify each mention as describing a deployed system, an intention, legal risk language, governance content, or a residual category, and validate the classification against 250 manually labelled paragraphs (κ = 0.66 on the binary deployment measure, against an intra-rater bound of κ = 0.78). Three results follow. First, the increase is driven by firms that began disclosing AI after 2022, and these firms describe little deployment: 23.3% describe at least one working system, against 74.9% among earlier adopters, with deployment accounting for 5.2% of their AI text against 27.3%. Second, among firms already disclosing AI before 2022, deployment paragraphs are flat between 2021 and 2025 while risk language grows fivefold and governance language grows from zero. Third, no text-based measure — neither the count of mentions nor the deployment share — is significantly associated with return on assets or sales growth in any of nine specifications. Corporate disclosure became a weaker signal of AI adoption after 2022: the substantive component did not disappear, but was diluted by language that grew several times faster.