AI use and risk disclosure by investment advisers

Abstract

Introduction Investment advisers increasingly report using artificial intelligence (AI), but little is known about what they disclose about it in fiduciary filings. The SEC brought its first AI-washing enforcement actions in 2024, so regulators and researchers need a reproducible measure of AI disclosure. We provide large-sample evidence on AI disclosure in Form ADV Part 2A brochures and show how the estimate depends on measurement design. Methods We drew a stratified random sample of 400 SEC-registered investment advisers from a population of 16,223 firms, crossing adviser type (private-fund vs. wealth/retail) with 2024 assets-under-management (AUM) quartile. A large language model (LLM) classified the 388 available brochures under a prespecified five-label typology: investment-process use, operational or client-service use, AI-related risk, explicit non-use, and named AI vendors. We treated the LLM as a measurement instrument and validated it three ways: same-family reproducibility, blinded cross-family recoding of 180 brochures, and human verification of 158 cases. Secondary analyses compared brochures with 283 matched firm websites and applied a screen derived from SEC enforcement orders. Results AI use is disclosed in 23.7% of brochures (22.7% survey-weighted), while 27.6% contain AI-related risk language. The within-brochure difference is significant (McNemar p = 0.036). Disclosure rises from 10.4% in the smallest AUM quartile to 43.8% in the largest, and is higher among private-fund advisers. Against human coding, the composite any-use measure achieves precision of 0.76 and sensitivity of 0.81. The same corpus yields estimates of 9.9%, 13.1%, and 23.7% under alternative constructs, adviser universes, and weighting schemes. A literal keyword rule misses or mislabels roughly a quarter of cases. A raw 17.3-percentage-point brochure–website gap disappears after conditioning on document length. None of the five enforcement-screen matches was adjudicated as an unsubstantiated promotional claim. Discussion Advisers present AI at least as often as a risk as a capability, and disclosure is concentrated among larger and private-fund advisers. AI-disclosure prevalence reflects measurement design as much as adviser behavior, so estimates should be reported with their construct, universe, weighting, and validation. The validated framework and an open reproducibility capsule provide a baseline for monitoring AI claims and AI-washing in investment management.