# Same Question, Smaller Answer: The Hidden Bias in AI Money Advice

> Source: <https://smarterarticles.co.uk/same-question-smaller-answer-the-hidden-bias-in-ai-money-advice?pk_campaign=rss-feed>
> Published: 2026-11-14 02:00:30+00:00

## Same Question, Smaller Answer: The Hidden Bias in AI Money Advice

Before GPT-4o offered a single word of guidance, it had already decided the woman was poorer.

That is the quiet finding buried in the tables of a benchmark posted to arXiv on 28 October 2026. Its author, Kriti Mahajan, asked OpenAI's model for financial advice on behalf of synthetic Indian personas, and had it estimate, alongside that advice, what each persona owned and earned before and after following it. The request itself never changed. Only the labels did: gender, caste, education, employment, age, whether the person lived in a city or a village, and in one large set of prompts, which Indian district they called home. Nobody told the model how much money anyone had. It filled that in for itself.

Across the gender-only prompts, profiles labelled male were assigned initial total resources (the sum of wealth, income and savings) 60.1 per cent higher than otherwise comparable female profiles. Then came the advice. After it, the model projected that the men would still hold 46.8 per cent more than the women. In the prompts that also carried a caste label, the post-advice gap between men and women was 49.4 per cent. The paper's regressions control for age, employment, education, residence and caste or gender, and the male coefficient stays positive and statistically significant in both the gender-only and the caste-augmented specifications.

It would be easy to dismiss this as one model, one country, one researcher. But it arrived in a year when economists at MIT Sloan and Stanford reached an uncomfortably similar conclusion about American users, when the UK's financial regulator found many young investors think chatbot output is regulated, and when Vanguard warned that general-purpose AI tools sit outside the rules that bind human advisers. What happens when advice that is slightly worse for some people, depending on who they are, is followed for forty years? And who, exactly, is responsible for it?

## A Poorer Woman Before the First Word of Advice

The Indian Financial Automated Advice Bias benchmark, or IFAAB, is a counterfactual audit: hold everything constant except the attribute under test, and differences in output can be attributed to that attribute. The dataset contains 31,734 GPT-4o responses, split into three experiments: 1,404 gender-only prompts, 7,020 gender-by-caste prompts and 23,310 gender-by-district prompts.

What the model produced for each persona was a block of advice plus six estimated figures: wealth, income and savings before the advice, and the same three after it. There is no ground truth here: every number is the model's own projection. That limitation is also the point. The audit is not measuring what Indian women own. It is measuring what GPT-4o assumes they own, and what it believes its own advice will do for them.

In the gender-only experiment, the average male profile was assigned initial resources of ₹264,700 against ₹165,400 for the average female profile. After advice, those figures rose to ₹386,900 and ₹263,600. One subtlety cuts against the simplest version of the story. In percentage terms the women's projected resources grew faster, by roughly 59 per cent against roughly 46 per cent for men, which is why the relative gap narrows from 60.1 to 46.8 per cent. In rupees, though, the gap widens, from ₹99,300 before advice to ₹123,300 after it. The model is not simply making women poorer. It starts them poorer, gives them advice it thinks will help, and still projects them ending up further behind in absolute terms.

The texture of the advice helps explain why. Mahajan classified the topics the model raised and found that men more often received guidance on investing, formal employment, networking and entrepreneurship, while women more often received guidance on saving, self-help groups, home-based work and basic literacy. The paper's own paired examples are blunt. A male profile is told: “Start a small monthly SIP in a balanced mutual fund.” A comparable female profile is told: “Consider investing in a recurring deposit or post office savings scheme.” Another female profile is advised to “Save a small, consistent amount each month, even if it is just Rs. 100.”

None of these suggestions is reckless, which is what makes the pattern so hard to see. No single answer looks discriminatory. The bias lives in the distribution: in who gets steered towards growth assets and who gets steered towards safety, across thousands of conversations.

## A Caste Gradient in the Machine

The caste experiment adds a second axis. Relative to profiles labelled General category, the model's projected post-advice total resources were lower by ₹72,600 for Other Backward Classes profiles, ₹114,000 for Scheduled Caste profiles and ₹113,100 for Scheduled Tribe profiles, all statistically significant. Profiles with no caste specified were indistinguishable from General ones. The projected gains from the advice itself were also smaller for OBC, SC and ST profiles.

A model that merely mirrored existing inequality in its starting estimates would be troubling enough. A model whose advice is projected to deliver smaller gains to people already assigned less is doing something closer to amplification. Mahajan's abstract makes the general point directly, warning that LLM-based financial agents can reproduce and reinforce existing social inequalities.

The district experiment complicates matters in an instructive way. When the prompts included the persona's district, the aggregate gender gap shrank dramatically: the initial resource gap fell to 2.5 per cent and the post-advice gap to 7.9 per cent. Mahajan reads this as evidence that, without geographic context, the model leans on gender as a proxy for financial capacity. Give it a richer description and it relies less on the crude signal. But the gap did not vanish, men still received higher projected post-advice income, and the paper notes that only 30 prompts sat behind each district, so local estimates are noisy.

The study has real limits, and its author lists them. Only GPT-4o was tested, the gender-by-caste personas were all fixed at age 25 for budget reasons, and the paper warns that its results should not be read as universal properties of LLM financial advice. Mahajan notes that the work was done independently and self-funded. It is a preprint, not yet peer reviewed.

Still, its direction matches a growing body of work. In 2024, Alejandro Salinas, Amit Haim and Julian Nyarko of Stanford audited leading language models, GPT-4 among them, by asking for advice about named individuals across 42 prompt templates. They found the advice “systematically disadvantages names that are commonly associated with racial minorities and women,” and that “names associated with Black women receive the least advantageous outcomes.” A 2026 Centre for Economic Policy Research discussion paper by Richard Foltyn and Jonna Olsson, cited in the study below, reported an average gender gap of 1.8 percentage points in recommended equity allocations across 33 language models.

## Prompts Written by Real People

The most rigorous attempt to measure what this kind of bias costs over a lifetime comes from Taha Choukhmane, an assistant professor of finance at MIT Sloan, Tim de Silva of Stanford Graduate School of Business, and two MIT Sloan colleagues, Weidong Lin and Matthew Akuzawa. Their paper, “AI Financial Advice: Supply, Demand, and Life Cycle Implications,” appeared as NBER Working Paper 35574 in August 2026 and on arXiv the same month. It won a Swiss Finance Institute Outstanding Paper Award.

Their innovation was to stop writing prompts themselves. They recruited a demographically balanced sample of 1,000 US adults through the research platform Prolific and asked each to write three prompts: a description of their financial situation, a request for advice on spending versus saving, and a request for investment advice. Those real, messy, human questions were fed to large language models, principally GPT-5.2, and the advice was then run through a life-cycle simulation in which virtual people aged 22 onwards followed it year after year, through job losses, market slumps and death.

The headline result is genuinely good news. Following the chatbot's advice would move most people closer to what economists consider sound behaviour: broader participation in diversified equity funds, equity shares that decline with age, larger savings buffers. “We were somewhat surprised by how good the advice was,” Choukhmane told MIT Sloan's Ideas Made to Matter. De Silva, speaking to Stanford GSB Insights, put it this way: “AI seems to be nudging people in the right direction.”

The bad news is in the heterogeneity. The advice was not equally good for everyone, and the differences followed familiar lines. Prompts written by women generated recommendations that left simulated wealth at age 60 about $60,000, or 4.1 per cent, lower on average than prompts written by men, in 2025 dollars. Prompts from people who answered at least one basic financial literacy question incorrectly produced wealth about $46,000, or 4.1 per cent, lower. Prompts from people who had never used AI for financial advice produced wealth about $100,000, or 5.7 per cent, lower than prompts from experienced users.

This is where the frequently quoted “4 to 6 per cent” comes from, but those are three separate comparisons between different pairs of groups, not a single range. The gender and literacy gaps were driven mainly by portfolio choice: women were recommended an average equity share between ages 22 and 60 that was 2.94 percentage points lower than men's. The AI-experience gap was driven almost entirely by saving advice instead. Different mechanisms, similar magnitudes, all compounding in the same direction.

“Regular people are not writing their prompts the way a finance professor is,” Choukhmane said. His colleague was blunter: “The way you write the questions matters a ton.”

## Separating What You Ask From Who You Are

If the story ended there, it would be about prompting skill, and the remedy would be education. But a second experiment changes the moral.

They took prompts from respondents that did not mention gender and randomly attached a label, such as “I am a man” or “I am a woman”, to the front. That allowed them to split any gender difference into two parts. The demand side captures differences in what men and women actually write. The supply side captures what the model does when the identical prompt arrives under a different label.

For the recommended diversified equity share, moving from a prompt written by a man and labelled male to one written by a woman and labelled female reduced the recommendation by 1.50 percentage points in a given period. About two-thirds of that, 0.96 points, was demand-driven. The remaining third, 0.54 points, was supply-driven. Holding every word of the question fixed, telling the model the asker was a woman made it recommend less in equities.

Two caveats are essential. First, the one-third figure applies to the per-period equity recommendation, not to the cumulative $60,000 wealth gap. The authors say explicitly that these coefficients “do not cumulate advice differences over the life cycle”. It is reasonable to infer that some meaningful share of the lifetime gender gap is supply-driven, but the paper does not state that one-third of the wealth gap is. Second, the race results looked different. Differences between Black and White respondents, and between Asian and White respondents, were smaller and almost entirely demand-driven, with a supply effect close to zero for race.

The authors offer two readings of the gender label effect. The model may be using gender as a statistical proxy for things that correlate with it, such as risk preferences or labour-market risk. Or it may be reproducing stereotypes from its training data. Even under the charitable reading, they write, “gender is a coarse signal, and this type of inference would ideally be made explicit to users.” On race, they suggest model providers may police race more stringently than gender during alignment, but warn that an explicit statement such as “I am Black” is exactly the signal such safeguards are built to catch. Names, dialect and context are subtler, and the Stanford names study suggests that is where trouble lives.

The supply-demand distinction is the hinge of this whole debate. Demand-side differences can be framed, with some justice, as the consequence of what people choose to ask. Supply-side differences cannot. They are a property of the product. When a human adviser gives a woman more conservative advice than a man with an identical file, that is the sort of thing a regulator, an ombudsman or a court can examine. When a chatbot does it, it is not obvious that anyone has jurisdiction to look.

## The Arithmetic of a Small Tilt

Why should a gap of a few percentage points matter? Because compounding converts small persistent differences into large terminal ones.

A simple illustration, using assumptions chosen for clarity rather than drawn from any study, shows the mechanism. Suppose a portfolio earns 5 per cent a year, and suppose that holding about three percentage points less in equities, roughly the lifetime gender difference the MIT and Stanford researchers found, costs about 0.15 percentage points of annual expected return. That is what you would get if equities outperformed safer assets by 5 points a year. A pound invested at 22 and left alone until 60 ends up about 5.3 per cent smaller under the cautious tilt. A steady stream of equal annual contributions ends up about 3.3 per cent smaller. After ten years the lump-sum cost is barely visible, around 1.4 per cent. It is the decades that do the damage.

Those illustrative figures sit, reassuringly or alarmingly, in the same neighbourhood as the 4.1 per cent the full simulation produced for women. The simulation also holds income, shocks and market returns fixed across groups, so it isolates the effect of the advice alone. Whatever other disadvantages a particular user already carries into retirement, the advice gap is added on top of them, not subtracted from them.

Apply the same lens to the Indian evidence. If a model begins by assuming a woman, or a Scheduled Caste household, has less, and tailors its advice to that assumption by suggesting smaller sums and safer products, then the advice becomes a self-fulfilling forecast. Each recommendation shapes the next year's balance. Each balance shapes the next question. A user who returns to the same chatbot every year, as the MIT and Stanford simulation assumes, is not getting one biased answer. She is getting a biased trajectory.

There is a further compounding effect that no simulation captures: who uses these tools in the first place. “A lot of the people who would benefit from financial advice are precisely the people who don't have a lot of resources,” Choukhmane told MIT Sloan. Stanford's write-up of the same paper noted that around 40 per cent of the human prompt writers had less than $10,000 in savings. These are the people least likely to pay a human adviser, most likely to rely on a free chatbot, and least able to absorb a slow drain on returns. The demographic discount is applied hardest to those with the least room for error.

## Counting the People Who Lost Money

How much harm is already visible? The evidence is thinner than the headlines suggest, and some of it has been counted twice.

The best-documented figures come from NerdWallet. In an online survey conducted by The Harris Poll on 23 and 24 June 2026 among 2,003 US adults, 26 per cent said they had used an AI chatbot to find answers to personal finance questions. Among those who had acted on a chatbot's advice, 29 per cent said it hurt their financial situation, 39 per cent said it helped and 32 per cent said it made no difference. The harm figure rests on a smaller base of 496 chatbot users, is self-reported, and cannot say whether the advice, the market or the user was to blame.

Vanguard's research report “The AI advice frontier: Use, trust, and the human edge”, dated September 2026 and written by Andy Reed, Nathan Young, Xiao Xu, Samantha Ostrom and Fiona Greig, is often quoted as finding that three in ten people who act on AI advice lose money. The report does contain that line. But its own survey, fielded in June 2026 among 6,686 Vanguard personal investors, is not its source. Vanguard attributes it to the NerdWallet study, by Erin El Issa. They are one finding, reported twice, not two independent confirmations.

Vanguard's own survey is more revealing. About 32 per cent of its investors had used AI for some form of financial guidance, rising to 43 per cent among Gen Z and millennial investors. Many did not understand the tools' legal status. Asked whether general-purpose AI has a legal obligation to act in their best financial interest, according to the report's chart, 34 per cent of Gen Z and millennial users, 20 per cent of Gen X users and 13 per cent of baby boomers and older users said they were not sure, and a further small share, between 3 and 4 per cent in each group, believed it did.

The UK picture is starker. On 27 August 2026 the Financial Conduct Authority published research from a survey of 666 UK adults aged 18 to 40 who held or were considering investments, run through the Attest platform on 24 July. Four in five less experienced investors had used AI to help with investing. Fifty-six per cent trusted AI tools. And 44 per cent wrongly believed AI-generated financial information was regulated. Thirty-eight per cent thought it acceptable to base an investment decision solely on AI output, and 32 per cent wrongly believed they would be compensated by the Financial Services Compensation Scheme or the Financial Ombudsman Service if AI advice went wrong.

Read alongside the bias research, these numbers describe an unusual situation: a substantial minority of users believe they are protected, and they are not.

## What a Human Adviser Owes You

In the UK, giving a personal recommendation on investments is a regulated activity. Firms that do it must be authorised by the FCA and must assess suitability under the FCA's conduct of business rules, gathering enough information about a client's circumstances, objectives and capacity for loss to justify the recommendation. Since 31 July 2023 they have also been bound by the Consumer Duty, which requires firms to act to deliver good outcomes for retail customers, to act in good faith, to avoid causing foreseeable harm and to enable and support customers in pursuing their financial objectives. Crucially for this debate, the Duty's monitoring rules require firms to consider whether any group of retail customers is experiencing different outcomes. A UK firm whose advice tool systematically recommended lower equity exposure to women on identical facts would, at least in principle, have a problem it was obliged to find.

The UK has also just widened the space between full advice and generic guidance. Under the FCA's targeted support regime, set out in policy statement PS25/22 and in force since 6 April 2026, authorised firms can offer suggestions designed for groups of consumers with shared characteristics, for pensions and investments, without conducting a full individual suitability assessment. Targeted support is lighter than advice, but it is still a regulated activity, with permission, conduct rules and the Consumer Duty attached.

In the United States, the Investment Advisers Act of 1940 imposes a federal fiduciary duty on registered investment advisers. In June 2019 the Securities and Exchange Commission published an interpretation clarifying the fiduciary duty an investment adviser owes its clients. On the same day it adopted Regulation Best Interest, which requires broker-dealers recommending securities to retail customers to act in the customer's best interest and not to place their own interests ahead of the customer's, with compliance required from 30 June 2020. Critics, including Commissioner Robert Jackson in dissent, argued that Reg BI fell short of a true fiduciary standard. But it is enforceable, and it attaches to a person.

In India, the Securities and Exchange Board of India's Investment Advisers Regulations, 2013 require anyone acting or holding themselves out as an investment adviser to register. SEBI amended them in December 2024, adding a duty for registered advisers using AI tools to disclose the extent of that use to clients and making the adviser solely responsible for the security, confidentiality and accuracy of client data processed through such tools. In August 2024, SEBI also amended its rules to bar regulated entities from associating with unregistered individuals who offer investment advice or make performance claims, a measure aimed squarely at so-called finfluencers.

Across three legal systems, the principle is the same. When a person or firm tells you what to do with your money, they owe you something: suitability, care, loyalty, disclosure, a route to complain and, often, a compensation scheme behind them.

## Why the Chatbot Owes You Nothing

General-purpose chatbots fall outside almost all of this, and not by accident. Each regime has a perimeter, and each perimeter was drawn for a world in which advice was either personal and paid for, or published and impersonal.

The FCA has said so directly. Its August 2026 release states that “General purpose AI chatbots are not regulated,” while noting that tools specifically set up to provide financial advice would be likely to fall within its remit. The distinction is coherent from a regulator's point of view: a general-purpose assistant is not set up as an advice service. But it means little to a 25-year-old who types a question about her pension into the same app she uses for everything else, receives a tailored, confident answer and follows it. Lucy Castledine, the FCA's director of consumer investments, framed the regulator's message to consumers as a matter of self-reliance: “you need to understand how you're protected and continue to use your own judgement.”

The US perimeter has its own famous gap. In Lowe v SEC in 1985, the Supreme Court held that the Investment Advisers Act excludes “the publisher of any bona fide newspaper, news magazine or business or financial publication of general and regular circulation,” and read the Act's legislative history as aimed at personalised advice rather than impersonal publishing. A chatbot generating output tailored to an individual's circumstances is an odd fit for a carve-out designed for newsletters, but how it fits the Act remains an open question. Meanwhile the SEC's 2023 proposal to govern conflicts of interest arising from predictive data analytics, which would have applied to advisers and brokers using such tools, was formally withdrawn on 12 June 2025 along with 13 other pending proposals. It would not have reached unregistered chatbot providers in any case.

India's definition of an investment adviser turns on providing advice for consideration, and the regulations exclude advice given through newspapers, magazines or “any electronic or broadcasting or telecommunications medium, which is widely available to the public.” A free chatbot available to anyone with a phone sits uncomfortably close to that exclusion. The December 2024 amendments address AI only when a registered adviser uses it. They do not touch the AI provider whose tool a household consults directly.

Vanguard's report puts the asymmetry crisply. Investment advisers “have long been strictly regulated, are generally held to a fiduciary standard, and are accountable to their clients,” it says, while “general-purpose AI tools have not been subject to any of the same regulations,” even when they offer explicit investment advice. The consequence is that the very finding at the heart of this story, a supply-side bias in which the model treats identical requests differently according to gender, falls into a space no conduct regulator currently supervises. A human adviser who did it would face suitability rules, outcome monitoring and complaints. The chatbot faces, at most, general consumer and equality law, which was not drafted with this kind of harm in mind.

## Closing the Perimeter Gap

Regulators are not blind to this. The FCA's Mills Review, led by executive director Sheldon Mills and published on 6 July 2026, concluded that retail finance is moving towards services that are “AI-enabled, continuous and delegated.” Its first recommendation to the FCA Board was to “Secure and adapt the regulatory perimeter.” Its seventh was to develop a trusted public-interest AI-enabled financial capability service. Law firm commentators at Freeths noted that it did not recommend a standalone AI regulatory regime for retail financial services. The direction of travel is visible. The destination is not. Several options follow from the evidence.

The first is functional regulation. Vanguard argues for “applying appropriate investor-protection standards to functionally equivalent activities, regardless of whether guidance is provided by a human advisor or an AI system.” If a system takes a person's age, income and goals and produces a specific portfolio recommendation, it is doing what an adviser does. The difficulty lies in drawing the line without sweeping in every search engine and explanatory answer, or pushing general-purpose providers simply to refuse all money questions, which would leave the people Choukhmane worries about with nothing.

The second is mandatory bias testing. The IFAAB benchmark and the MIT and Stanford label-swap method both show that counterfactual audits are cheap, repeatable and revealing. Swap a gender or caste label, hold everything else constant, measure the difference. Requiring providers whose models answer large volumes of financial questions to run and publish such audits, in the way the Consumer Duty already expects authorised firms to monitor group outcomes, would turn an invisible bias into a reportable metric. It would also make the supply-side share, the part of the gap that is unambiguously the model's doing, something a regulator could track over time.

The third is disclosure at the point of use. The MIT and Stanford authors argue that if a model is inferring something from gender, that inference “would ideally be made explicit to users.” The FCA and Vanguard findings suggest many users also need to be told, at the moment they ask, that the answer carries no regulatory protection and no compensation. Vanguard's own report cautions that education “is necessary but not sufficient”, adding that “Investors shouldn't have to bear the burden of discerning high-quality versus high-accessibility advice.”

The fourth is to build a better public option. The Mills Review's idea of a public-interest financial capability service, and the UK's new targeted support regime, point towards a world in which accessible, AI-assisted guidance exists inside the perimeter, with the duties attached, rather than outside it. If that option is good enough and easy enough to reach, it competes directly with the unregulated chatbot for the users who most need help.

None of these is a complete answer. But the evidence now justifies at least one modest step everywhere: treating systematic differences in AI-generated advice by gender, caste or other protected characteristics as a conduct problem rather than a curiosity.

## The Advice Gap That Advice Creates

So, is the most consequential wealth gap of the AI era the one created by the advice itself?

The evidence does not quite support that strong claim. The MIT and Stanford study found that, for most people, following chatbot advice would improve on what they already do. Lower-income households with little savings may gain more from an imperfect free adviser than they lose to its biases. Two-thirds of the measured gender gap in equity advice came from what people asked, not from the label the model saw. Pay, inheritance and housing shape wealth gaps too, and the paper measures only one channel among many. And the IFAAB results come from one model, as a preprint, with no ground truth beneath the numbers.

But the weaker version of the claim is strong enough to worry about. The audits show that leading models, given identical requests, tilt their advice according to gender, and in GPT-4o's case in India according to caste, in directions that line up with existing disadvantage. The simulation shows that tilt, compounded across a working life, can be worth tens of thousands of dollars. The surveys show that a quarter of Americans have already asked a chatbot about their money, that many cannot tell good guidance from bad, and that a large share of young UK investors believe they are protected when they are not. And the legal analysis shows that the supply-side portion of the bias, the part that is plainly the product's own doing, sits in the one place that human advisers' fiduciary, suitability and outcome-monitoring duties do not reach.

What makes this gap distinctive is not its size. It is its invisibility. A woman who gets a more cautious answer than a man would have got will never see the man's answer. A woman told to save Rs. 100 a month never sees the monthly mutual fund plan that, in the Indian audit, went to men. No single conversation looks wrong, so there is nothing obvious to complain about. The discount is applied silently, in small amounts, every time the question is asked, and it compounds.

Human advisers were bound by duties because societies decided that people telling others what to do with their savings could not be left to self-interest and good intentions. The machines now doing some of that telling were never brought inside those duties. Until they are, the most reliable protection against the demographic discount is the one the FCA recommends, and the one its own research suggests many people are not using: your own judgement, applied to advice you have no way of knowing was different for someone else.

## References and Sources

1. Mahajan, Kriti. “The Indian Financial Automated Advice Bias (IFAAB) Dataset: Auditing Socioeconomic and Demographic Disparities in LLM-Generated Financial Advice in India.” arXiv preprint 2610.38239v1, submitted 28 October 2026. [https://arxiv.org/abs/2610.38239v1](https://arxiv.org/abs/2610.38239v1) (full text:[https://arxiv.org/html/2610.38239v1](https://arxiv.org/html/2610.38239v1) )
2. Choukhmane, Taha (MIT Sloan School of Management and NBER); de Silva, Tim (Stanford Graduate School of Business); Lin, Weidong (MIT Sloan); Akuzawa, Matthew (MIT Sloan). “AI Financial Advice: Supply, Demand, and Life Cycle Implications.” NBER Working Paper 35574, August 2026. DOI: 10.3386/w35574. [https://www.nber.org/papers/w35574](https://www.nber.org/papers/w35574)
3. Vereckey, Betsy. “AI financial advice is surprisingly good, especially if you ask the right questions.” MIT Sloan School of Management, Ideas Made to Matter, 21 July 2026. [https://mitsloan.mit.edu/ideas-made-to-matter/ai-financial-advice-surprisingly-good-especially-if-you-ask-right-questions](https://mitsloan.mit.edu/ideas-made-to-matter/ai-financial-advice-surprisingly-good-especially-if-you-ask-right-questions)
4. Blumberg, Deborah Lynn. “What AI Tells People Seeking Low-Cost Financial Advice.” Stanford Graduate School of Business, Insights, 13 July 2026. [https://www.gsb.stanford.edu/insights/what-ai-tells-people-seeking-low-cost-financial-advice](https://www.gsb.stanford.edu/insights/what-ai-tells-people-seeking-low-cost-financial-advice)
5. MIT Sloan School of Management. “CFI Paper Highlight: AI Financial Advice, Supply, Demand, and Life Cycle Implications.” 2026. [https://mitsloan.mit.edu/centers-initiatives/cfi/ai-financial-advice-supply-demand-and-life-cycle-implications](https://mitsloan.mit.edu/centers-initiatives/cfi/ai-financial-advice-supply-demand-and-life-cycle-implications)
6. Reed, Andy; Young, Nathan; Xu, Xiao; Ostrom, Samantha; Greig, Fiona. “The AI advice frontier: Use, trust, and the human edge.” Vanguard Research, September 2026 (Vanguard AI & Advice Survey, June 2026, n=6,686). [https://corporate.vanguard.com/content/dam/corp/research/pdf/the_ai_advice_frontier.pdf](https://corporate.vanguard.com/content/dam/corp/research/pdf/the_ai_advice_frontier.pdf)
7. El Issa, Erin. “Data: Americans Are Using Chatbots for Financial Advice. Risky or Rewarding?” NerdWallet, 2026 (survey conducted by The Harris Poll, 23 to 24 June 2026, n=2,003 US adults). [https://www.nerdwallet.com/finance/studies/using-ai-for-personal-finances](https://www.nerdwallet.com/finance/studies/using-ai-for-personal-finances)
8. Financial Conduct Authority. “Young investors trust AI more than TV or celebrities.” Press release, 27 August 2026. [https://www.fca.org.uk/news/press-releases/young-investors-trust-ai-more-tv-or-celebrities](https://www.fca.org.uk/news/press-releases/young-investors-trust-ai-more-tv-or-celebrities)
9. Financial Conduct Authority. “AI and the future of retail financial services (The Mills Review).” 6 July 2026. [https://www.fca.org.uk/publications/corporate-documents/mills-review](https://www.fca.org.uk/publications/corporate-documents/mills-review)
10. Financial Conduct Authority. FCA Handbook, PRIN 2A (Consumer Duty), including PRIN 2A.9 (monitoring of consumer outcomes). [https://handbook.fca.org.uk/handbook/PRIN/2A/](https://handbook.fca.org.uk/handbook/PRIN/2A/)
11. Financial Conduct Authority. “PS25/22: Supporting consumers' pensions and investment decisions: rules for targeted support.” 2025. [https://www.fca.org.uk/publications/policy-statements/ps25-22-supporting-consumers-pensions-investment-decisions-rules-targeted-support](https://www.fca.org.uk/publications/policy-statements/ps25-22-supporting-consumers-pensions-investment-decisions-rules-targeted-support)
12. Penningtons Manches Cooper. “Rethinking investor engagement: targeted support and its regulatory implications.” April 2026. [https://www.penningtonslaw.com/insights/rethinking-investor-engagement-targeted-support-and-its-regulatory-implications/](https://www.penningtonslaw.com/insights/rethinking-investor-engagement-targeted-support-and-its-regulatory-implications/)
13. Sidley Austin LLP. “SEC Adopts Regulation Best Interest and Form CRS; Issues Investment Advisers Act Interpretations.” 13 June 2019. Archived copy: [https://web.archive.org/web/20200805080948/https://www.sidley.com/en/insights/newsupdates/2019/06/sec-adopts-best-interest-standard-of-care-for-broker-dealers-clarifies-fiduciary-duty](https://web.archive.org/web/20200805080948/https://www.sidley.com/en/insights/newsupdates/2019/06/sec-adopts-best-interest-standard-of-care-for-broker-dealers-clarifies-fiduciary-duty)
14. Cleary Gottlieb Steen & Hamilton. “SEC Adopts Best Interest Standard for Broker-Dealers and Fiduciary Duty Guidance for Investment Advisers.” June 2019. [https://clearygottlieb.com/news-and-insights/publication-listing/sec-adopts-best-interest-standard-for-broker-dealers-and-fiduciary-duty-guidance](https://clearygottlieb.com/news-and-insights/publication-listing/sec-adopts-best-interest-standard-for-broker-dealers-and-fiduciary-duty-guidance)
15. US Securities and Exchange Commission. “Conflicts of Interest Associated with the Use of Predictive Data Analytics by Broker-Dealers and Investment Advisers” (File No. S7-12-23), proposal and withdrawal notice. [https://www.sec.gov/rules/2023/07/s7-12-23](https://www.sec.gov/rules/2023/07/s7-12-23)
16. Paul Hastings LLP. “SEC Withdraws 14 Pending Rule Proposals.” June 2025. [https://www.paulhastings.com/insights/client-alerts/sec-withdraws-14-pending-rule-proposals](https://www.paulhastings.com/insights/client-alerts/sec-withdraws-14-pending-rule-proposals)
17. Supreme Court of the United States. Lowe v. Securities and Exchange Commission, 472 U.S. 181 (1985). Legal Information Institute, Cornell Law School. [https://www.law.cornell.edu/supremecourt/text/472/181](https://www.law.cornell.edu/supremecourt/text/472/181)
18. Securities and Exchange Board of India. SEBI (Investment Advisers) Regulations, 2013 (text via Institute of Cost Accountants of India). [https://icmai.in/upload/pd/SEBI-IA-Regulations-2013.pdf](https://icmai.in/upload/pd/SEBI-IA-Regulations-2013.pdf)
19. Taxmann. “SEBI Mandates Investment Advisers to Disclose Use of 'AI Tools' to Clients in Providing Investment Advice.” [https://www.taxmann.com/post/blog/sebi-mandates-investment-advisers-to-disclose-use-of-ai-tools-to-clients-in-providing-investment-advice](https://www.taxmann.com/post/blog/sebi-mandates-investment-advisers-to-disclose-use-of-ai-tools-to-clients-in-providing-investment-advice)
20. AZB & Partners. “Overhaul of the Regulatory Framework for Investment Advisers and Research Analysts.” 3 February 2025. [https://www.azbpartners.com/bank/overhaul-of-the-regulatory-framework-for-investment-advisers-and-research-analysts/](https://www.azbpartners.com/bank/overhaul-of-the-regulatory-framework-for-investment-advisers-and-research-analysts/)
21. Business Today. “Sebi tweaks regulations for finfluencers, removes content of over 15,000 unregulated entities.” 31 August 2024. [https://www.businesstoday.in/personal-finance/investment/story/sebi-tweaks-regulations-for-finfluencers-removes-content-of-over-15000-unregulated-entities-443789-2024-08-31](https://www.businesstoday.in/personal-finance/investment/story/sebi-tweaks-regulations-for-finfluencers-removes-content-of-over-15000-unregulated-entities-443789-2024-08-31)
22. Salinas, Alejandro; Haim, Amit; Nyarko, Julian. “What's in a Name? Auditing Large Language Models for Race and Gender Bias.” arXiv preprint 2402.14875, February 2024, revised January 2025. [https://arxiv.org/abs/2402.14875](https://arxiv.org/abs/2402.14875)
23. Foltyn, Richard; Olsson, Jonna. “The worth of a 'wo': Gender bias in financial advice from LLMs.” Centre for Economic Policy Research Discussion Paper 21323, 2026. [https://cepr.org/publications/DP21323](https://cepr.org/publications/DP21323)
24. Freeths LLP. “FCA's Mills Review: Agentic AI, Consumer Duty and the future of financial services.” 2026. [https://www.freeths.co.uk/insights-events/legal-articles/2026/fca-s-mills-review-agentic-ai-consumer-duty-and-the-future-of-financial-services/](https://www.freeths.co.uk/insights-events/legal-articles/2026/fca-s-mills-review-agentic-ai-consumer-duty-and-the-future-of-financial-services/)
25. Financial Conduct Authority. FCA Handbook, COBS 9 (Suitability, including basic advice). [https://www.handbook.fca.org.uk/handbook/COBS/9/](https://www.handbook.fca.org.uk/handbook/COBS/9/)

**Tim Green**
*UK-based Systems Theorist & Independent Technology Writer*

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at [smarterarticles.co.uk](https://smarterarticles.co.uk), challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

**ORCID:** [0009-0002-0156-9795](https://orcid.org/0009-0002-0156-9795)
**Email:** [tim@smarterarticles.co.uk](mailto:tim@smarterarticles.co.uk)

Listen to the free weekly [SmarterArticles Podcast](https://www.smarterarticles.fm)