Investment Advisers Reveal AI as Risk More Often Than Capability

Artificial intelligence has swept through the financial industry faster than almost any technology before it, and now the first large-scale audit of what investment advisers actually say about AI in their official filings has produced a strikingly counterintuitive result. A study published in Frontiers in Artificial Intelligence analyzed nearly 400 fiduciary disclosure documents filed with the U.S. Securities and Exchange Commission and found that advisers are more likely to describe AI as a risk to be warned about than as a capability they possess. In an era when industry surveys suggest that roughly 63 percent of advisory firms are using or piloting AI, only about 24 percent of formal brochures disclose any AI use at all, while nearly 28 percent contain language flagging AI as a source of investment, operational, cybersecurity, or regulatory risk.

The research, conducted by independent researchers Samir Chincholikar and Robin Chawla, arrives at a politically charged moment. In March 2024, the SEC brought its first enforcement actions for so-called AI-washing, the practice of exaggerating or fabricating AI capabilities to attract clients. The agency alleged that Delphia (USA) Inc. had claimed machine-learning capabilities it never developed, and that Global Predictions Inc. promoted itself as the first regulated AI financial advisor with AI-driven forecasts that were never substantiated as advertised. The firms settled for civil penalties of $225,000 and $175,000 respectively. Those cases made one thing clear to regulators and researchers alike: before anyone can detect false AI claims, someone needs a reliable way to measure whether firms mention AI in the first place.

That measurement problem is harder than it sounds, and the study treats it as the centerpiece of its design. The authors built their sample from the SEC’s Form ADV Part 1 data, identifying 16,223 active SEC-registered advisers reporting regulatory assets under management at the end of 2024. From that population they drew a stratified random sample of 400 firms, crossing two business models, private-fund advisers versus wealth and retail advisers, with four size quartiles, and then retrieved the narrative Part 2A brochures, the formal disclosure documents describing each firm’s practices, strategies, conflicts, and risks. Usable text was obtained for 388 brochures.

Instead of trusting a single automated classifier, the researchers deployed a large language model, GPT-4o with a fixed model snapshot and temperature set to zero, as a measurement instrument under a prespecified five-label typology frozen before any results were calculated. The labels distinguished AI in the investment process, such as research, portfolio construction, and trading; AI in operations or client service; AI as a disclosed risk factor; explicit statements of non-use; and references to named AI vendors or products. Each positive classification required verbatim supporting quotes from the brochure text.

The validation architecture is where the study distinguishes itself methodologically. The authors ran three separate checks. First, they assessed reproducibility within the same model family by reclassifying 60 brochures with GPT-4o-mini using the same frozen rubric. Second, they had 180 brochures blindly recoded by Anthropic’s Claude, a different model family entirely, with the validation statistics corrected for a two-phase sampling design using Horvitz-Thompson inverse-probability weights. Third, they performed targeted human verification on 158 cases, coded by two domain-knowledgeable coders who were blind to the machine labels. Against human coding, the composite any-use measure achieved a precision of 0.76, sensitivity of 0.81, and specificity of 0.94, meaning the classifier caught most genuine disclosures while making a meaningful share of false positive calls, concentrated mainly at the boundary between investment-process AI and operational AI.

The headline findings are stark. Disclosed AI use appeared in 23.7 percent of brochures, or 22.7 percent when survey-weighted to the full population. But risk language appeared in 27.6 percent, and the difference held within the very same documents: a paired analysis showed 77 brochures contained both risk language and disclosed use, while 30 contained risk language alone against only 15 with use alone. The McNemar test confirmed the within-brochure gap was statistically significant. In other words, when AI enters a fiduciary filing, it enters more often as a hazard to disclose than as a capability to advertise, a pattern consistent with the brochure’s institutional function of cataloging material risks rather than promoting technology.

Disclosure was also radically uneven across the industry. Among private-fund advisers, prevalence climbed from 12 percent in the smallest assets-under-management quartile to 56 percent in the largest; among wealth and retail advisers the gradient ran from roughly 9 percent to 30.4 percent. A logistic regression confirmed that both the size gradient and the private-fund effect held after controlling for each other, with the ordered trend tests strongly significant across the combined sample. The pattern suggests that firms with systematic investment approaches, alternative data pipelines, and specialized technology infrastructure have more disclosure-relevant AI references, although the cross-sectional design cannot separate the underlying mechanisms or establish causation.

Perhaps the study’s most consequential finding is that no single true prevalence number exists. The same corpus of 388 brochures yielded estimates of 9.9 percent, 13.1 percent, and 23.7 percent depending on whether the construct was narrowed to the literal phrase artificial intelligence, restricted to the wealth and retail universe, or broadened to the full classifier-based any-use measure. A crude keyword rule achieved only 0.75 precision and 0.74 recall against the classifier, meaning a simple phrase search both misses and mislabels roughly a quarter of cases. The authors argue that any AI-disclosure estimate should be reported alongside its construct definition, adviser universe, weighting scheme, and validation statistics, because the number is a joint product of the measurement design as much as of the filings themselves.

The study also compared fiduciary brochures against the public-facing websites of 283 matched firms and found a raw 17.3 percentage point gap, with 23 percent of brochures disclosing AI use against just 5.7 percent of websites, and 58 firms disclosing only in the brochure versus only nine disclosing only on the web. Yet the authors refuse the tempting interpretation that advisers are more candid in regulatory filings than in marketing. Brochures in the matched sample contained 7.3 times the text of the website corpora, and once document length was included in a statistical model, the venue effect vanished entirely, with length itself strongly predictive of detection. Per 100,000 characters, websites actually produced slightly more qualifying passages than brochures. It is a cautionary lesson for anyone comparing disclosure across documents of unequal length.

On the question of AI-washing itself, the enforcement-anchored screen, built from seven textual claim patterns distilled from the two SEC orders, flagged only five of 388 brochures. Two proved to be false-positive textual matches, and the other three contained contextual or risk language that qualified the AI references. None was adjudicated as an unsubstantiated promotional claim resembling the charged conduct. The authors emphasize that the screen captures only language patterns from two enforcement actions and cannot estimate the true prevalence of AI-washing, but the result suggests that the charged style of exaggerated AI claims was rare in formal fiduciary filings. The full framework, including the frozen rubric, prompts, model specifications, validation records, and a deterministic Code Ocean capsule that regenerates every reported number, is openly available, providing regulators and researchers with a validated baseline for tracking how AI claims and AI risks migrate through investment management’s disclosure ecosystem.

Subject of Research: AI use and risk disclosure in SEC investment adviser filings

Article Title: AI use and risk disclosure by investment advisers

Article References: Chincholikar, S., & Chawla, R. (2026). AI use and risk disclosure by investment advisers. Frontiers in Artificial Intelligence, 9, Article 1969257. https://doi.org/10.3389/frai.2026.1969257

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1969257

Keywords: AI-washing, investment advisers, Form ADV, Securities and Exchange Commission, large language models, text classification, fiduciary disclosure, machine learning, financial regulation, survey weighting, text as data, private funds