How much do investment advisers disclose about using AI?
A peer-reviewed study found affirmative AI-use language in 23.7% of 388 US adviser brochures and risk language in 27.6%. It measures formal disclosure—not underlying adoption, safe performance or AI-washing.
The short version
3 key points
- 1.Researchers sampled 400 advisers from 16,223 SEC-registered firms with reported assets under management and classified 388 available Form ADV Part 2A brochures.
- 2.Affirmative AI-use disclosure appeared in 23.7% of brochures, while 27.6% contained AI-related risk language; the measures can overlap and neither establishes actual adoption or misconduct.
- 3.Against human coding, the composite any-use classifier had precision of 0.76, sensitivity of 0.81 and specificity of 0.94, leaving meaningful measurement error.
The finding: roughly a quarter disclosed AI use, but that is not an adoption rate
A peer-reviewed study of US investment advisers found affirmative AI-use language in 23.7% of 388 available fiduciary brochures. AI-related risk language appeared in 27.6%. Those figures show what firms wrote in Form ADV Part 2A, not whether they had deployed AI, whether a system worked safely or whether a claim was misleading.
That boundary is the most important result for investors and regulators. An adviser may use an internal or experimental tool that is not material to the brochure, while another may describe algorithmic methods without using the label AI. Risk language can refer to market, vendor, cybersecurity or model risks without saying the adviser itself uses AI. Absence of a phrase is therefore not evidence of non-use, and presence is not an audit of the technology.
The researchers deliberately treated the language model as a measurement instrument rather than ground truth. They showed that the estimated prevalence changes when the definition of use, the adviser universe and the weighting change. The headline 23.7% is a carefully tested textual measure, but it is still a constructed measure.[1][2]
How the sample was built
The team began with SEC Form ADV Part 1 bulk data. Of 16,709 active SEC-registered advisers at the end of 2024, 16,223 reported regulatory assets under management and formed the sampling frame. The authors drew a stratified random sample of 400 firms, crossing adviser type—private-fund or wealth and retail—with four asset-size quartiles.
Current Part 2A brochures were available and classifiable for 388 firms. The study used a prespecified five-label scheme: AI in the investment process; AI in operations or client service; AI-related risk; explicit non-use; and named AI vendors or products. A composite 'any-use' measure combined the functional-use labels and vendor references, with a vendor-excluded alternative reported because vendor naming was harder to classify reliably.
Stratification matters because a simple count could overrepresent particular firm types or asset bands. The unweighted any-use estimate was 23.7%; the survey-weighted estimate was 22.7%. Disclosure rose from 10.4% in the smallest asset quartile to 43.8% in the largest, and was higher among private-fund advisers. Those gradients describe brochure language within the 2024 sampling frame, not future behaviour or all financial firms.[1][2]
The classifier was checked, but it was not error-free
A large language model classified the brochures, and the authors tested it in three ways: same-family reproducibility, blinded recoding of 180 brochures with a different model family, and targeted human verification. Human reviewers assessed 120 brochure cases within a total of 158 verified cases across the study's document types. The verification estimates were weighted for the two-phase selection design.
Against human coding, the composite any-use measure achieved precision of 0.76, sensitivity of 0.81 and specificity of 0.94. In plain language, some passages labelled as AI use were not confirmed by people, and some human-confirmed uses were missed. Risk language was easier to identify: its reported precision was 0.92, recall 0.99 and F1 score 0.95. Investment-process and operational categories were less consistent.
The human check also inherited an upstream limitation. Reviewers examined the keyword-selected passages shown to the model rather than every word of each brochure. It therefore tested whether retrieved passages received the right label, but not whether the retrieval step found every relevant passage. The authors disclose this directly, which is essential for interpreting the reported sensitivity.[1][2]
Risk language was common, but it does not prove safer practice
Among the 388 brochures, 77 contained both affirmative-use and risk language, 30 contained risk language only, 15 disclosed use only, and 266 contained neither. The within-document difference between risk and use was statistically significant in the authors' paired test. This suggests advisers often frame AI as something to warn clients about as well as a capability to describe.
A disclosure is still not a control. The study does not test whether a firm validates models, protects client data, monitors vendors, prevents unsuitable recommendations or gives staff effective override authority. Nor does it show that longer risk sections improve understanding. Investors need concrete information about where AI enters a decision, what human review exists, what data are used and what happens when the system is wrong.
Larger advisers disclosed more often, but the paper cannot determine whether they use more AI, face more scrutiny, have more sophisticated legal drafting or simply produce longer brochures. Asset size and adviser type are associated with the measured language; they are not causal explanations.[1]
The study does not estimate AI-washing
The researchers created a secondary screen based on two SEC enforcement orders concerning misleading AI claims. Five brochures triggered that textual screen. After review, none was adjudicated as an unsubstantiated promotional claim of the kind alleged in those orders. The authors correctly refuse to turn five screening matches into an AI-washing prevalence estimate.
That restraint matters. A phrase resembling language from an enforcement case is only a prompt for investigation. Establishing misconduct requires evidence about what the firm represented, what capabilities actually existed, materiality, context and the applicable legal standard. Automated screening can prioritise records; it cannot issue a finding.
The study also compared 283 firms observed in both brochures and public websites. AI use appeared in 23.0% of brochures and 5.7% of sites, but the raw 17.3-point gap disappeared after accounting for document length. Because brochure and website lengths barely overlapped, the paper does not claim firms are more candid in one venue.[1][2]
What this changes for investors, advisers and regulators
For investors, the practical lesson is to ask a narrower question than 'Does my adviser use AI?' A useful answer identifies the task, the decision rights retained by people, the data and vendor dependencies, the error and monitoring process, and whether an AI output can directly change a portfolio or client communication. Generic capability and risk language cannot answer those questions.
For advisers and compliance teams, the five-part typology offers a way to review consistency across brochures, marketing and operational reality. Firms should not infer that mentioning risk cures an exaggerated capability claim. They should connect each material statement to evidence, ownership and review, and keep disclosures current as vendors and workflows change.
For regulators, the open classification framework could support trend monitoring across filings, provided its error rates remain visible and flagged documents receive human review. The method is more defensible than a literal keyword search, which the study says missed or mislabelled roughly a quarter of cases. It is not a substitute for examination, technical testing or enforcement evidence.[1]
What evidence would change the assessment
The next study should match disclosure to independently verified firm-level use. A representative sample could document which tools were deployed, which tasks they performed, whether outputs reached clients or portfolios, and what controls operated. Comparing those facts with contemporaneous brochures would measure true uptake–disclosure alignment rather than infer it from unrelated surveys.
A stronger text measure would add full-document human review for a random subset, publish retrieval as well as classification error, and test performance across filing years and newly emerging terminology. Repeating the analysis prospectively would show whether firms add precise descriptions or merely longer generic risk clauses after regulatory attention.
Evidence that clear, specific disclosures improve client understanding or lead to better oversight would support treating the language as a meaningful protection. Evidence of persistent gaps between actual deployment and formal statements, or misleading claims confirmed through examinations, would justify a more concerning assessment. For now, the paper establishes a reproducible baseline for disclosure—not a verdict on the industry.[1][2]
What this means for people
- Clients need to know where AI affects advice, communications or portfolio decisions and which person remains accountable for errors.
- Compliance teams can use automated screening to prioritise review, but model flags require human interpretation and evidence about actual systems.
- Smaller advisers disclosed AI less often in this sample, but the study cannot tell whether that reflects lower use, shorter documents or different drafting practices.
Global context
The sample covers US SEC-registered investment advisers and 2024 regulatory filings. Other countries use different disclosure rules, adviser populations and enforcement systems, so the percentages should not be exported without a new sampling frame and locally validated classifier.
What the evidence does not yet show
- The outcome is brochure language, not verified AI adoption, system safety, investment performance or misconduct.
- Only 388 of 400 sampled advisers had classifiable brochures, and all measures refer to a 2024 US regulatory population frame.
- The any-use classifier had precision of 0.76 and sensitivity of 0.81 against human coding, leaving material false positives and false negatives.
- Human reviewers judged retrieved passages rather than complete brochures, so retrieval failures were not fully measured.
- The study received no declared financial support; the authors report aggregate, pseudonymised results and withhold the adviser crosswalk.
What to watch next
- Matched audits comparing actual adviser AI deployments with contemporaneous brochure language
- Full-document human validation and year-to-year tests of retrieval and classification error
- More specific disclosures about tasks, vendors, human review, data, failure modes and client recourse
- Confirmed regulatory findings rather than automated screening matches when estimating AI-washing
Living evidence record
Impact record IAI-15MBEE5
Evidence stage
Studied
Confidence
Supported
Reporting basis
Source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
14 October 2026
Source trail
2 direct sources across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Related-source reporting disclosure
This record analyses 2 linked source records around the same underlying development. The extra records add method, date or context, but they do not by themselves constitute independent replication of every performance claim or predicted outcome.
Evidence trail
Sources used for this report
Links checked 14 October 2026
Research question, topics and regions
How often US investment advisers describe AI use and AI-related risk in formal brochures, and how classification choices change the estimate
Topics: Investment advice · AI disclosure · AI-washing · Regulation · Financial services · Text classification · Fiduciary duties
Regions: United States · North America · Global
This report is labelled source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
The Impact Brief · Free
Keep following the stories that matter to you.
Get our weekday briefing on what changed and why it matters. Choose your topics below.
Continue the story
There’s more to this story
Finance & Business
What does Character.ai’s bankruptcy mean for users?
Character.ai says its service and paid subscriptions continue during a court-supervised sale, but the platform could be wound down if no buyer agrees to keep it operating. Users can export data now; the proposed Disney licence is not a completed sale.
7 min · 2 sources
Finance & Business
Is the AI boom propping up global growth—or concentrating it?
The IMF says AI investment and its hardware supply chain are helping activity withstand other shocks, but it does not assume an extra AI productivity gain in its baseline. Its 13 October outlook projects 3.1% global growth in 2026 while warning that the benefit is concentrated among economies connected to the technology cycle.
8 min · 3 sources
Finance & Business
Who is financing AI's power backbone?
General Catalyst, Koch Equity Development and co-investors have agreed to put $2 billion of preferred equity into Axiom at a $37.5 billion enterprise value. The capital is expensive, the separation is unfinished and the disclosed terms do not prove AI demand will deliver the promised returns.
8 min · 2 sources
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.