Investment advisers mention AI in nearly a quarter of their fiduciary brochures—but they mention AI-related risks even more often. The surprising part is that the headline percentage changes substantially depending on how disclosure is measured. Investment advisers increasingly report using artificial intelligence, but little is known about what they disclose about it in fiduciary filings.

The SEC brought its first AI-washing enforcement actions in 2024, creating a need for a reproducible measure of AI disclosure. Artificial intelligence can enter an advisory business through investment research, portfolio construction, trading, risk management, client service, operational processes, or third-party technology.

These uses may raise different disclosure considerations, so a useful measure must establish whether firms mention AI at all, what kind of use they disclose, and whether AI is presented as a capability or a risk. This study takes a measurement-first approach to AI disclosure among U.S.

investment advisers and asks how frequently advisers disclose AI use across assets-under-management and adviser-type strata. It also asks whether AI appears more often as a claimed capability or as a disclosed risk, and how fiduciary brochures compare with public-facing websites.

An estimate of AI disclosure depends on what counts as AI use, which documents and advisers are observed, and how reliably the underlying text can be classified. The large language model is treated as a measurement instrument rather than ground truth, with classifications evaluated through same-family reproducibility, blinded cross-family recoding, and targeted human verification.

The study identifies 16,223 active SEC-registered advisers reporting regulatory assets under management and draws a stratified random sample of 400 firms across adviser type and AUM quartile. Current Form ADV Part 2A brochures were categorized for 388 firms. A prespecified five-label typology distinguishes AI use in the investment process, AI use in operations or client service, AI-related risk disclosures, explicit statements of non-use, and named AI vendors or products.

The classification rubric and model specification were specified before outcome analysis, followed by same-family reproducibility and blinded, independent cross-family recoding of 180 brochures. In March 2024, the SEC announced settled enforcement actions against two investment advisers for false or misleading statements about their purported use of AI.

The SEC order alleges that Delphia did not develop the capabilities it claimed to use machine learning to analyze aggregate client data in a predictive algorithmic model. Global Predictions represented itself as the first regulated AI financial advisor and promoted AI-driven forecasts that the SEC found were not substantiated as advertised.

The two companies agreed to civil penalties of 225,000 U.S. dollars and 175,000 U.S. dollars, respectively. The sampling frame was built from the SEC’s Form ADV Part 1 bulk filing data.

The study merged 27,679 unique Central Registration Depository identifiers from filings between 2011 and 2024 and kept advisers active as of December 31, 2024. Of 16,709 active SEC-registered investment advisers, 16,223 reported regulatory assets under management and formed the population frame for sampling.

A five-label classification framework was created and locked in before outcome analysis to distinguish substantively different forms of AI-related disclosure. The 388 brochures were classified using GPT-four-o as the main large language model.

A fixed model snapshot was used with the temperature set to zero to minimize avoidable generation variability. Each brochure was scored using the prespecified rubric, and the model had to output binary labels with supporting text for positive classifications.

A separate validation exercise used an independently operated model from another provider and model family. Anthropic Claude recoded 180 brochures using the same frozen definitions without access to the GPT-four-o classifications, so the recoding was blind to the primary labels.

Agreement between two model families does not by itself establish that either model is measuring the intended construct, so the study conducted targeted human verification on 158 cases. The verification frame included all 60 brochures with any model disagreement, a stratified sample of 60 brochures with agreement, and 38 website cases.

Two domain-knowledgeable coders were blind to the model labels and used the same prespecified rubric; each case received one human code. Against the human labels, the composite any-use measure returns sensitivity of zero point eight one, specificity of zero point nine four, and precision.

Of the 388 observed brochures, 23.7 percent contain disclosure of at least one form of AI use. Looking down the prevalence column, risk-factor disclosures appear in 27.6 percent of brochures, compared with 18.3 percent mentioning AI in the investment process.

The broad “any mention” row reaches 31.7 percent, while explicit prohibition or non-use is only 2.1 percent. Disclosures related to risks associated with AI occurred more often than affirmative disclosures about AI use, with 27.6 percent of brochures referencing AI, machine learning, or related technology as a source of risk.

AI appears in investment-adviser brochures as a disclosed technological capability, but more often as something to be warned about. A risk-factor reference does not imply that the adviser itself uses AI, so risk disclosure is reported separately from the any-use measure.

Of the 388 brochures, 77 carry both risk language and disclosed use, 30 carry risk language only, 15 disclosed use only, and 266 carry neither. The within-brochure difference is 3.9 percentage points, with a paired bootstrap 95 percent confidence interval from 0.5 to 7.2 and an exact McNemar p value of 0.036.

Disclosed use of AI increased substantially across the 2024 AUM strata. Among private-fund advisers, prevalence increased from 12 percent in the lowest AUM quartile to 24 percent, 34 percent, and 56 percent in the second, third, and highest quartiles.

For wealth and retail advisers, the corresponding estimates were 8.7 percent, 8.3 percent, 14.6 percent, and 30.4 percent. The pattern is not limited to one adviser category. Looking across each quartile pair, the solid blue private-fund bars remain visibly taller than the hatched wealth-and-retail bars, with the separation becoming especially pronounced in the rightmost group.

The chart also makes the concentration of disclosures at the largest asset-under-management quartile clear for both adviser types. The independent cross-family validation provided a more demanding test. Taking the verification-sampling design into consideration, the composite any-use measure obtained precision of zero point seven six, recall of zero point eight four, an F1 score of zero point eight zero, overall agreement of 90.8 percent, and a Cohen’s kappa of zero point seven four.

The main classifier identified most of the independently identified positive cases and showed substantial agreement with the blinded cross-family recoding. Looking across the indicator rows, the risk-factor label has an F1 of zero point nine five one, while explicit non-use has an F1 of zero point six zero nine.

The named-vendor row shows the widest precision–recall contrast: zero point five two four precision alongside one point zero zero zero recall. Current website observations were obtained and usable for 283 sampled advisers.

Within this matched sample, disclosed AI appeared in 23.0 percent of Form ADV Part 2A brochures but only 5.7 percent of current principal-site marketing materials. The paired bootstrap 95 percent confidence interval for the 17.3-point difference is 12.0 to 22.6.

The figure’s two bars make the venue contrast immediately visible: the fiduciary-brochure bar reaches just above twenty percent, while website marketing remains near six percent. The caption adds an important qualification: matched brochures contain seven point three times as much text, and controlling for document length removes the venue difference.

The study measures disclosed AI use rather than underlying adoption, so the estimates are properties of documents, not of the technologies firms deploy. Disclosed AI use appears in roughly a quarter of U.S. investment-adviser fiduciary brochures, and it is unevenly distributed.

Prevalence rises steeply across the 2024 AUM strata and is higher among private-fund than wealth and retail advisers. AI-related risk language is at least as common as affirmative use language within the same documents, so AI enters these filings as a risk to be disclosed as much as a capability to be described.

The headline figure depends materially on how the construct is defined, which advisers are counted, and how much text each document contains. The same corpus supports estimates from roughly 10 to 24 percent, and a binary indicator rises with document length.

Prevalence is therefore a property of measurement design as much as of the filings from which it is drawn. The study’s central lesson is that AI disclosure is both uneven and measurement-sensitive: larger and private-fund advisers disclose more, while risk language is at least as common as capability language.

The number is a property of the measurement design as much as the documents.