AI in Investment Advice: Risk Warnings Outnumber Claims of Use in Client Brochures

The findings show how AI language varies across advisers, how automated readers can misunderstand financial documents, and why a percentage describing AI disclosure needs an explanation of how it was calculated.

Artificial intelligence can help an investment adviser research markets, handle paperwork or answer clients, making a simple question surprisingly difficult to answer: what does a firm actually tell people about its use of the technology? A review of U.S. adviser brochures found that roughly one in four disclosed AI use, with warnings about its risks appearing even more often than descriptions of its capabilities.

The research, "AI use and risk disclosure by investment advisers," by Samir Chincholikar and Robin Chawla, published in Frontiers in Artificial Intelligence, examines what these disclosures can reveal about financial firms. The findings show how AI language varies across advisers, how automated readers can misunderstand financial documents, and why a percentage describing AI disclosure needs an explanation of how it was calculated.

How the Researchers Read the Fine Print

The researchers randomly selected 400 advisers from a population of 16,223 SEC-registered firms, giving equal representation to private-fund and wealth or retail advisers across four groups based on assets under management. They obtained usable Form ADV Part 2A brochures for 388 firms; these documents describe services, investment strategies, conflicts and material risks. Eleven sampled firms were exempt from filing a brochure, and one document could not be converted into usable text, so missing brochures were not automatically counted as having no AI disclosure.

GPT-4o classified the documents using rules fixed before the results were calculated, with supporting excerpts required for positive labels. Five overlapping categories separated AI in investment decisions, AI in operations or client service, AI-related risks, explicit statements of non-use, and named AI vendors or products. This distinction matters because a warning about AI disrupting markets does not establish that the adviser uses it, and mentioning a product provides weaker evidence than describing an actual activity.

Risk Warnings Lead, with Larger Firms Disclosing More

AI-use disclosure appeared in 23.7% of brochures, falling slightly to 22.7% after weighting the sample to reflect the population of brochure-filing advisers. Risk language appeared in 27.6%, producing a 3.9-percentage-point difference within the same documents. Investment-process AI featured in 18.3%, operational or client-service use in 6.7%, named vendors in 5.4%, and explicit non-use in 2.1%. The categories overlapped, with 31.7% containing at least one kind of AI-related mention.

Disclosure rose from 10.4% in the lowest asset-management group to 43.8% in the highest. Private-fund advisers disclosed use in 31.5% of brochures, compared with 15.4% among wealth and retail advisers. Within private funds, the smallest-to-largest group comparison was 12% versus 56%; for wealth and retail advisers, it was 8.7% versus 30.4%. The size pattern remained after accounting for adviser type, establishing an association that the research could not explain in terms of cause and effect.

The paper places these findings beside an industry survey reporting that about 63% of advisers were using or piloting AI. The two figures describe different populations, dates and activities, so their gap cannot show how many firms are concealing AI use. Administrative experiments may never appear in a client brochure, and some brochures describe machine-learning methods that survey respondents might not label AI. Silence in a filing is not proof that the technology is absent.

Why the Way AI Is Counted Changes the Answer

A narrow measure based on the literal phrase "artificial intelligence" produced a population-weighted estimate of 9.9% for wealth and retail advisers. Broader classification in that same population gave 13.1%, compared with 23.7% across the full study sample. Against the model classifier, the keyword rule missed roughly a quarter of qualifying cases and incorrectly included about a quarter of its matches. Dropping vendor-only references reduced overall disclosure to 22.4%, leaving the main patterns intact.

The researchers checked the automated reading through repeat classification of 60 brochures with GPT-4o-mini, blinded recoding of 180 with Claude, and human verification of 158 cases comprising 120 brochures and 38 websites. Against human judgments, the main brochure-use measure had 76% precision and 81% sensitivity, meaning about three-quarters of its positive classifications were confirmed and it detected about four-fifths of human-identified positives. Independent model recoding gave a disclosure estimate of about 21.6%, close to the original result.

Separating investment use from operational use proved harder than identifying risk language, making the broader use measure more dependable than its narrower categories. Human reviewers assessed the same keyword-selected passages shown to the models, so validation could not reveal relevant material that retrieval missed elsewhere in a brochure.

Firm categories and asset groups came from year-end 2024 data, with documents collected in 2026; checks using older asset data preserved the size pattern without fully accounting for subsequent changes in firms. The analysis captures a snapshot rather than changes in disclosure over time.

What Websites and AI-Washing Checks Can Really Reveal

For 283 firms with usable website content, AI-use disclosure appeared in 23.0% of brochures and 5.7% of websites. Brochures contained 7.3 times as much text, and the apparent difference disappeared after accounting for document length. The length ranges barely overlapped, making it difficult to separate document type from text volume. Website classification detected only about a third of human-identified use cases, adding uncertainty; the findings cannot establish that advisers are more candid in brochures than in marketing.

A separate screen used claim patterns drawn from the SEC's 2024 enforcement actions against Delphia and Global Predictions, cases involving misleading descriptions of AI capabilities. It flagged five brochures for review: two were false matches, and three contained AI statements supported or qualified by context or risk warnings. None was judged an unsubstantiated promotional claim resembling those enforcement cases. The screen's rate of missed cases was unknown, so the findings cannot estimate how widespread AI-washing is.

The study covers U.S. SEC-registered advisers and current principal websites, excluding wider marketing such as social media, advertisements and press releases. Its public reproducibility package allows others to rerun the analysis, with anonymised adviser identities limiting complete reconstruction of the original document collection. The authors propose tracking brochures over time, matching disclosures with actual firm-level AI adoption, and examining more communication channels to understand when technological claims enter public view.