AI financial advice needs someone responsible when it is wrong

When I worked with predictive models at Wells Fargo, there were people around the model who had to understand what it was doing and live with the decision that followed. That experience comes back to me now that chatbots are giving people recommendations about savings, debt and retirement accounts.

The Associated Press covered a new Gallup survey about financial guidance on August 7. Roughly one in five Americans who sought financial advice in the prior year used AI. Among U.S. adults overall, about three in ten had at least some confidence in AI's expertise for managing money. Only 3 percent said they had a great deal of confidence in it.¹

The Gallup category is broader than a bank or brokerage product. It can mean asking the same general-purpose chatbot used for a work question or a dinner recipe. A separate 2026 study by MIT and Stanford researchers gives us a view of that behavior: among 455 Prolific participants who said they had used AI for financial advice, 71.9 percent named ChatGPT, 17.4 percent Gemini and 2.2 percent Claude. Because those users came from an online panel, I use the product split only as a direct description of that sample. The paper did not separate paid subscriptions from free accounts.²

Professional advice is expensive, and plenty of people are starting with whatever answer they can afford. A chatbot is available late at night, does not charge for the first question and can explain an index fund without making someone feel foolish for asking. Gallup also found that 73 percent of Americans who sought guidance relied on their own internet research, while 32 percent used a professional financial adviser.³

The access gap grows for younger adults. About a quarter of Gen Z and millennial advice seekers used AI, while only 14 percent of Gen Z seekers and 21 percent of millennials used a professional adviser. Most younger respondents never reached a professional adviser at all.

Personal financial recommendations carry more responsibility than explanations

I use AI to work through unfamiliar subjects all the time. It can be very good at explaining a term, showing the steps in a calculation or giving me questions to ask someone who knows more than I do, and I find that useful.

Once a chatbot moves from explaining a Roth IRA to recommending what someone should do with one, it needs facts about the life around the account: a job that could end next month, a parent who needs care or an emergency fund that has already been spent. The person asking may not know which facts belong in the prompt, and the chatbot has no duty to go looking for what was left out.

MIT professor Taha Choukhmane told the AP that AI can be useful for definitions and explanations that help someone make better use of other sources. Certified financial planner Bobbi Rebell described the missing obligation: the AI is not a fiduciary and does not know the person's whole life.¹

The United States Securities and Exchange Commission, FINRA and state securities regulators warn that AI-generated financial information may be inaccurate, incomplete, out of date or simply made up. Their investor alert tells people to confirm the underlying sources and check with a registered professional before acting.⁶

Confirming the source or paying a registered professional is much easier advice to follow when the professional is affordable. When that option is out of reach, I need the screen to mark where its knowledge ends and point the person toward help that is available.

Financial advice at a brokerage leaves more than a saved conversation

In my brokerage experience, financial advice never belonged to the adviser alone. The recommendation had to fit a customer record that included risk tolerance, time horizon, liquidity needs, income, other investments and the reason the customer was investing. Advisers maintained that information, supervisors reviewed the work, and legal, compliance and operations departments built much of the method around them.

I never mistook compliance for proof that a recommendation would be correct. Its value was the trail it created. The firm could see what the adviser knew about the customer, what was recommended, which disclosures were provided and who was responsible for the account.

Much of what I remember now appears in Regulation Best Interest. The SEC says a broker-dealer making a covered recommendation must consider the risks, rewards and costs in light of the customer's investment profile. SEC staff guidance lists income, assets, debts, goals, time horizon, liquidity needs, investing experience and risk tolerance among the information to consider. When enough information cannot be obtained, the staff says the firm generally should decline to make the account recommendation. The same guidance points to records of the customer information collected, the person responsible for the account and written communications about recommendations and advice.⁹

That leaves me wondering what sits behind a recommendation from a technology company. Has the customer's risk tolerance been asked, recorded and maintained? What happens when the customer changes jobs, spends the emergency fund or needs the money sooner than expected? A transcript can preserve what was typed without showing that the company had enough information to give the advice in the first place.

FINRA is asking those questions when its member firms use generative AI. Its 2026 oversight report describes formal review, comprehensive documentation, testing for reliability and accuracy, ongoing monitoring of prompts and outputs, model-version tracking and human review. FINRA's rules apply to its member firms, and the report presents several of those controls as current practices it observed rather than a checklist imposed on every firm. A general-purpose chatbot may sit outside even that boundary, which is why consumers need to know what kind of company is giving the recommendation.¹⁰

I also want to know how the model separates official filings and audited results from what I would call vibe research: social posts, recent headlines and the market's temporary view of a company. A list of links tells me which documents were retrieved. I still cannot see which source carried the most weight, whether a short-term opinion overpowered a decade of results or whether the model understood the difference.

Recent market concentration makes stock-picking FOMO easier

Anyone watching the market through 2023, 2024 and 2025 saw the same group of large technology companies return to the top of the conversation. At the end of 2025, the Magnificent Seven made up 34.9 percent of the S&P 500's market value and produced 42 percent of its return for the year. Across the prior three years, those seven companies contributed 55 percent of the index's total return.¹¹

A market-cap-weighted S&P 500 ETF participated in those gains because it owned the largest companies at large weights, giving its investor plenty of Magnificent Seven exposure. Someone holding an equal-weight index, smaller companies, an international fund or a personal portfolio could hear that "the market" was up and wonder why the account looked so different from the headline.

That gap can produce a powerful fear that everyone else found the obvious trade. By July 2026, State Street's analysis showed that the seven companies were no longer moving as a single group, even though they had driven a disproportionate share of returns in earlier years. Someone buying the label rather than studying the companies could already be acting on an old market story.¹²

Professional managers do not make stock picking look easy. S&P Dow Jones Indices found that 79 percent of active large-cap U.S. equity funds underperformed the S&P 500 in 2025. Its persistence report found that none of the top-quartile large-cap funds from 2021 remained in the top quartile through 2025.¹³ Those managers had research staffs, data and formal investment processes. A consumer asking a chatbot for a stock has much less help judging whether the recommendation found durable value or simply repeated the companies that had already won.

The SEC and FINRA have warned for years that social-sentiment tools can contain stale, incomplete or misleading information and can encourage impulsive decisions. Their advice is to compare that material with company disclosures, understand how the tool collects its data and keep short-term emotion from taking over a long-term plan.¹⁴ I would ask the same of an AI product before trusting a stock recommendation dressed in a polished explanation.

Corporate adoption of AI is growing while public confidence remains low

Gallup published two other findings this summer that belong beside the financial-guidance survey. Forty-seven percent of U.S. employees said their organization had integrated AI tools in the second quarter, up from 41 percent one quarter earlier. More than half said they now use AI in their own role at least occasionally.⁵

In a separate national survey, 27 percent of Americans said they trusted businesses at least some to use AI responsibly, down from 31 percent in 2025. Among adults ages 18 to 29, the share with at least some trust fell to 20 percent, while 41 percent said they had no trust at all.⁴

I would not subtract one percentage from another and call the remainder a trust score because the surveys asked different people about different decisions. I read them as evidence that AI is entering ordinary work and financial research quickly, while confidence in the companies deploying it remains low.

People judge financial advice by how it sounds and who appears to have written it

A preregistered study released August 10 tested what happened when researchers kept the substance of financial advice constant but changed its style and displayed source. Expert-style advice was rated more favorably than AI-style advice on nine of ten outcomes. Mislabeling the source could also change how people rated the AI advice.⁸

With 285 participants and only a first preprint, the study is too small and too new to carry a national conclusion. The question still feels important to me because people react to the words, the label and the confidence of the delivery along with the underlying recommendation. A polished response can create a feeling of competence before the person reading it has checked the facts or understood whose interests shaped the product.

Financial companies already know that presentation changes how advice is received. Their statements, disclosures and suitability questions are designed around the consequences of a recommendation. FINRA's 2026 oversight report leaves member firms' existing supervisory obligations in place when they use generative AI. Firms remain responsible for accuracy, communications, third parties, customer information and the people overseeing the technology.⁷

I would hold a consumer financial app to that standard even when it falls into a different legal category. The company chooses whether its guidance looks complete, personal or safe, and a disclaimer at the bottom cannot erase those choices.

Financial recommendations require current sources and customer context

An explanation of compound interest can be checked with arithmetic and current source material. Selling an investment requires decisions about taxes, timing, risk, other assets and what the money has to do next. An app giving both kinds of answers has to tell the customer when it has moved from an explanation into a recommendation.

Interest rates, tax limits and account rules change. I check those claims against a dated source before I act. When the model cannot show where a current number came from, the customer is left to guess whether the answer is still current.

Income, debt, time horizon, tax treatment and risk tolerance can change the recommendation. Put those inputs where the customer can see and change them. A wrong salary or an old contribution limit is much easier to catch before it becomes a transaction.

When a recommendation costs someone money, the customer needs to reach someone who can explain what happened, correct the record and say whether the institution stands behind its recommendation. Sending the customer back through the same chatbot after a loss would leave the company judging its own work.

Not every financial question requires an appointment. Definitions and calculations can stay inside the tool. When a recommendation could change what happens to someone's money, the customer still needs a clear way to reach someone who can take responsibility.

Our trust work looks at responsibility, verification, choice and recourse. Applied here, those are practical questions about who stands behind the recommendation, whether its facts and assumptions hold up, what control the customer keeps over financial information and where the customer can go after a bad result. Gallup did not test that framework, so this remains my reading of the problem rather than a finding attributed to respondents.

AI is already widening access to financial explanations, especially for people who may not be able to pay for professional advice. Gallup found people using it before they fully trust it. When a product can change what happens to someone's money, I still want a responsible person inside the company who can answer for what the product did.

Source notes

Adriana Morga, Associated Press, August 7, 2026, for AI use and confidence figures, differences by age, expert comments and survey methodology. Gallup and Edward Jones surveyed 5,075 U.S. adults ages 21 and older from March 20 through April 6 using Gallup's probability-based panel. The overall margin of sampling error was plus or minus 1.8 percentage points.
Taha Choukhmane, Tim de Silva, Weidong Lin and Matthew Akuzawa, AI Financial Advice, March 30, 2026, especially Appendix Figure A1, for the products named by 455 prior AI-for-finance users recruited through Prolific. The product split describes those respondents and is not a national estimate. The paper did not report whether they used paid or free versions.
Gallup, Fewer Than One in Five Financially Fulfilled in U.S., Canada, June 1, 2026, for the 73 percent internet-research and 32 percent professional-adviser figures. The U.S. results use the same 5,075-adult probability-panel study. Edward Jones sponsored the research; Gallup conducted it.
Gallup, Americans Cool Toward AI, July 28, 2026, for overall and young-adult trust in businesses' responsible use of AI. The separate Bentley University-Gallup study should not be merged statistically with the financial-guidance survey.
Gallup, Organizational AI Adoption Jumps Six Points, July 20, 2026, for employee-reported organizational adoption rising from 41 to 47 percent and individual workplace use reaching 52 percent. These are employee reports about workplace adoption, not consumer confidence or financial-advice usage.
U.S. Securities and Exchange Commission Office of Investor Education and Advocacy, NASAA and FINRA, Artificial Intelligence and Investment Fraud: Investor Alert, January 25, 2024, for warnings about inaccurate, incomplete, outdated or fabricated AI-generated information and the recommendation to verify sources and professional registration.
FINRA, GenAI: Continuing and Emerging Trends, 2026 Annual Regulatory Oversight Report, for the continuing application of supervision, communications, third-party, recordkeeping and customer-information obligations when member firms use generative AI. FINRA guidance applies to regulated member firms; I use it as evidence of institutional responsibility, not as a claim that every consumer chatbot is governed by the same rules.
Aryan Ramchandra Kapadia, Eshwar Chandrasekharan and Koustuv Saha, How People Evaluate AI-, Expert-, and Peer-Style Financial Advice, submitted August 10, 2026, a preregistered vignette experiment with 285 participants. The paper is a first-version preprint and has not been treated here as a nationally representative estimate or settled finding.
U.S. Securities and Exchange Commission staff, Standards of Conduct for Broker-Dealers and Investment Advisers Account Recommendations for Retail Investors, March 30, 2022, for the customer characteristics relevant to an investment profile, the need for sufficient customer information, and the cited broker-dealer and investment-adviser recordkeeping requirements. Regulation Best Interest governs covered broker-dealer recommendations to retail customers; an investment adviser's fiduciary duty arises under a separate framework. The bulletin also says Reg BI does not categorically require broker-dealers to document the basis for every recommendation, although the SEC encouraged firms to use a risk-based approach to that decision.
FINRA, GenAI: Continuing and Emerging Trends, 2026 Annual Regulatory Oversight Report, for governance, comprehensive documentation, reliability testing, prompt and output monitoring, model-version tracking and human review. These are FINRA observations and considerations for member firms, not universal legal requirements for every technology company.
S&P Dow Jones Indices, U.S. Equities Market Attributes, December 2025, for the Magnificent Seven's 34.9 percent year-end index weight, 42 percent contribution to the S&P 500's 2025 total return and 55 percent contribution across the three years ending in 2025.
State Street Global Advisors, Magnificent 7 No Longer Moving as One Trade, July 20, 2026, for the group's disproportionate contribution in 2023 and 2024 and the widening differences in company performance through June 2026. I use it as market analysis from an asset manager, not as a forecast or financial advice.
S&P Dow Jones Indices, SPIVA U.S. Year-End 2025, March 3, 2026, for the 79 percent underperformance rate among active large-cap U.S. equity funds in 2025, and U.S. Persistence Scorecard Year-End 2025, May 7, 2026, for the lack of five-year persistence among 2021's top-quartile large-cap funds. Fund-manager results do not directly estimate an individual investor's odds of selecting a winning stock.
SEC Office of Investor Education and Advocacy and FINRA, Social Sentiment Investing Tools: Think Twice Before Trading Based on Social Media, April 3, 2019, for risks from stale or misleading sentiment data, hidden agendas and impulsive trading, along with recommendations to review company disclosures and understand how the tool collects and analyzes information.

McQueen Analytics research note

McQueen Analytics' 2026 AI Trust work uses four U.S. adult online-panel surveys fielded from late May through June 12. The team is still reviewing claims and completing final population weighting, and financial guidance was outside the questionnaires. For this essay, I am using the wider trust research to ask who is responsible, what can be verified, what choices people keep and what happens after a failure. Those questions are my interpretation, not findings attributed to respondents.