Only 23.7% of investment advisers disclose AI use in their SEC brochures, and the number depends on how you count

Surveys keep telling us that most financial firms use AI. A study published on October 14, 2026 in Frontiers in Artificial Intelligence asks a different question: what do investment advisers actually write about AI in the documents they file with the SEC? The answer is less than the surveys suggest, and the study is careful to say why that is not a scandal.

What the researchers did

Chincholikar and Chawla drew a stratified random sample of 400 SEC-registered advisers from 16,223 firms, balanced across private-fund versus wealth/retail advisers and four size quartiles. They read the current Form ADV Part 2A brochure for 388 of them, the plain-language document that describes a firm's business, conflicts and risks. A large language model classified each brochure under a five-label scheme, and the authors checked it against a second model family and against human coding of 158 cases.

The headline numbers

AI use is disclosed in 23.7% of the brochures.⁶ Survey-weighting to the full adviser population gives 22.7%, so the sampling design is not what drives the result.⁷

AI shows up more often as a risk than as a capability: 27.6% of brochures mention AI as a risk factor.⁸ Within the same documents the difference is statistically significant (McNemar p = 0.036). In plain terms, advisers write about AI the way they write about other things that could go wrong.

Size matters a lot. Disclosure rises from 10.4% in the smallest asset quartile to 43.8% in the largest.⁹ Among private-fund advisers the top quartile reaches 56%, against 30.4% for the top wealth and retail quartile.¹⁰

Why the number is slippery

The most useful result is about measurement. Using the same corpus, the estimate runs from 9.9% to 23.7%, depending on whether you count only the literal phrase "artificial intelligence," which advisers you include, and how you weight them.¹¹ A simple keyword rule had precision of 0.75 and recall of 0.74 against the classifier, so it both misses and invents roughly a quarter of cases.

The classifier is not perfect either. Against human coders, the any-use measure had precision of 0.76 and sensitivity of 0.81. The model tended to read passages as AI use where a person would not, in 20 of 120 reviewed brochure cases against 1 the other way.

The gap with surveys

The authors cite a Schwab survey in which about 63% of advisers use or pilot AI, and note the roughly 39-point difference with 23.7%.¹² They warn against reading that as a nondisclosure rate. The groups, dates and definitions differ, and an adviser can use AI for internal tasks without any need to mention it in a client brochure.

The AI-washing screen

The study also screened for language resembling the two 2024 SEC AI-washing cases, Delphia and Global Predictions. Five of 388 brochures matched, and after review none was judged an unsubstantiated promotional claim.¹³ The authors say this is not an AI-washing prevalence estimate. The screen rests on two cases and a small number of positives.

What it shows, what it doesn't

It shows how often AI appears in formal filings, and how fragile any single prevalence number is. It does not show how many advisers use AI, whether disclosures are accurate, or whether any firm broke a rule. The size gradient is an association, with strata defined in 2024 and filings read in 2026.

The practical habit: when you see a statistic about how many firms use AI, ask what counted as use, who was counted, and where the claim was made. A brochure, a website and a survey answer are three different things. Our earlier checks of AI tool claims follow the same rule. This is analysis of a study, not advice about your money.