An AI investment adviser can sound knowledgeable about shares, bonds and savings, yet it may quietly imagine a customer who looks nothing like the person asking for help. Researchers at the University of California, Berkeley found that a model asked to simulate an ordinary investor tended towards the attitudes of younger, wealthier men unless demographic information was supplied. The work is a warning about assumptions inside AI-generated financial advice, not proof that every regulated adviser uses the same model or bias.
The Berkeley Haas research account describes an experiment comparing the investment preferences of real survey participants with responses generated using GPT-4. The model could reproduce meaningful differences between demographic groups when those groups were specified. Without that context, however, its default imagined investor was heavily skewed. A system can appear consistent and articulate while starting from the wrong idea of whom it is advising.
The AI investment adviser had a favourite sort of investor
The researchers surveyed more than 1,200 people about how they viewed stocks, bonds and cash. They then prompted GPT-4 to act as survey participants with different demographic backgrounds and compared the results. When age, gender and income were made explicit, the AI captured a number of broad patterns in people’s investment attitudes. Older respondents in the human survey, for example, tended to express greater preference for cash, while higher-income respondents were generally more comfortable with certain investments.
The researchers also examined how participants explained their choices. Both people and AI-generated responses discussed concepts such as risk, expected returns, experience and understanding. This suggested that the model was doing more than producing arbitrary rankings. It could imitate patterns of reasoning associated with the profiles it was given. That is promising for research tools, but it also raises the question of what happens when the profile is missing.
When the model was asked to respond as a general online survey participant without demographic instructions, the answers became less representative. Berkeley reports that 55 per cent of those responses resembled young, high-income men, despite that group making up less than 15 per cent of the comparison population. All of the model’s unprompted personas reported incomes above $54,000, and only 6 per cent described themselves as 39 or older. These are findings from the experiment’s simulated respondents, not a census of people using consumer finance chatbots.
That distinction is essential. The study did not show a commercial adviser automatically moving customers’ savings into unsuitable investments. It showed how one language model behaved when asked to portray investors, and how explicitly stating demographic characteristics altered its outputs. The potential danger is downstream: a financial product that assumes the wrong customer could frame risk and recommendations in an unhelpful way.
Why age, income and risk tolerance change the answer
Consider two people asking whether they should put more money into shares. One has stable income, ample savings and decades before needing the money. The other expects to draw on savings soon and cannot afford a large temporary fall in value. A generic response describing the long-term appeal of stock markets might be technically correct while being badly matched to one of those situations.
The difference is not simply whether somebody is young or old. An individual’s goals, income stability, debts, time horizon, existing investments and ability to tolerate losses all affect what advice would be appropriate. Age and income can provide context, but they are not substitutes for a full understanding of the person. A well-designed system should not rely on a stereotype in place of information that ought to be gathered directly.
This is where fluent AI output can be misleading. A model may confidently explain the advantages of diversification or the historical characteristics of asset classes. Those explanations may be broadly useful, but they are not necessarily tailored recommendations. If a person reads them as personal guidance without supplying essential information, the confident tone can create a false impression of suitability.
The problem is especially relevant to people who have limited access to traditional financial advice. Automated tools may promise low-cost guidance at any time, and that accessibility is attractive. It would be unfortunate if products intended to broaden access also carried invisible assumptions drawn from a narrow picture of the typical investor. Reducing that risk begins with acknowledging how the model responds when information is incomplete.
The experiment also revealed a possible strength
The Berkeley study was not simply an exercise in finding defects. Researchers also examined whether model responses were logically consistent when comparing investment options. If somebody prefers shares to bonds and bonds to cash, consistency would suggest they should prefer shares to cash. The model followed this type of consistent ranking more frequently than the human participants in the study.
Berkeley reports that GPT-4’s rankings met that consistency test 98.7 per cent of the time, compared with 84.4 per cent for human respondents. That does not mean the AI’s preferences were wiser or that people who express inconsistent preferences are making bad financial decisions. A person may be indifferent between choices, uncertain about risk or influenced by details the simple comparison does not capture. Logical consistency is useful evidence about the structure of an answer, not a guarantee of an appropriate portfolio.
A future advisory system might use this strength to identify contradictions in a customer’s stated priorities. For example, somebody may say they cannot tolerate losses yet also ask for a strategy built around maximum possible returns. A carefully designed tool could highlight the conflict and invite clarification. That would be a more defensible use of AI than confidently filling in the missing details with assumptions about the user’s finances.
A fairer system needs to ask better questions
The strongest practical implication is that an AI financial tool needs a sound process for finding out whom it is helping. It should not quietly decide the person’s age, income or appetite for risk because those details were omitted from a prompt. It should ask relevant questions, distinguish educational material from personal recommendations and make uncertainty visible when important information is missing.
That may sound obvious, but traditional financial services already devote considerable effort to collecting suitability information before offering certain kinds of advice. AI does not eliminate that responsibility. It creates a new interface through which information may be gathered, misunderstood or ignored. A pleasant conversation is not an adequate substitute for checks that protect a customer from taking a risk they cannot afford.
The authors have also made replication data and code available, allowing researchers to inspect how the experimental comparisons were produced. That transparency matters because debates about algorithmic bias can become vague. A clearly described sample, set of prompts and measurement approach makes it easier to ask whether the finding holds with another model, different demographic groups or a new version of the same system.
Future research would need to distinguish an AI system’s ability to simulate attitudes from its performance as a real adviser. A model can reproduce survey responses without understanding taxes, pension rules, investment charges or the regulatory duties attached to actual advice. Testing those aspects requires different evidence. The present work illuminates an assumption problem, not the overall suitability of a commercial platform.
What somebody using AI for money questions can do
The safest first use for an AI assistant is often educational. Ask it to explain the differences between shares, bonds and cash, to describe volatility or to show how fees affect a hypothetical investment over time. Then check important facts against authoritative documents and the terms of the actual product. Such tasks are different from asking a chatbot to decide what proportion of someone’s pension should be invested in a volatile asset.
For more personal questions, supplying relevant constraints can make the conversation less generic. That might include an investment time horizon, whether the money may be needed soon, the kinds of losses the person could tolerate and whether debt or emergency savings take priority. It does not mean sharing passwords, account numbers or unnecessary sensitive financial records in a chat. Specificity about the decision can be useful without handing a system unrestricted access to private accounts.
LiveAIWire has considered what the evidence says about AI investment managers, a broader question than the survey simulation studied at Berkeley. There are different categories of tools, including systems that explain concepts, models that generate suggestions and regulated services that make or implement investment choices. Users should establish which kind they are dealing with rather than assume that every product calling itself an AI adviser has the same oversight.
An earlier LiveAIWire discussion of whether AI financial advice is conservative or biased provides another perspective on the importance of model design and evaluation. A system may vary its tone, preferences or apparent caution depending on how a question is framed. Neither a reassuringly cautious answer nor an aggressive return forecast establishes that it is suitable for a particular investor.
The invisible assumption is the real risk
The most useful test of an AI financial assistant may be what it does when it lacks information. Does it ask an appropriate question? Does it explain that its response is general education? Does it acknowledge that different people can reasonably prefer different levels of risk? Or does it invent an imaginary investor and answer with polished certainty? The Berkeley results give a concrete reason to pay attention to that difference.
This is also a design issue for companies building financial tools. Fairness is not achieved simply by preventing an assistant from saying offensive things. It requires testing whether the system serves people with different financial circumstances and whether apparently sensible default assumptions systematically favour one group. Developers should measure what happens when users do not know which details the model needs.
LiveAIWire’s guide to using AI for stock-market research explores the wider question of getting useful analysis without confusing software output with investment certainty. Information can improve a decision, but risk belongs to the person making it. A tool should help reveal that risk rather than hide it behind a smooth response.
An AI adviser capable of modelling many kinds of investors is potentially valuable. Berkeley’s research suggests that capability must be deliberately brought into the conversation rather than assumed to appear by default. The dangerous answer is not always the obviously foolish one. Sometimes it is the elegantly written recommendation prepared for a person who was never actually there.
About the Author
Stuart Kerr is Technology Correspondent at LiveAIWire, covering artificial intelligence, cybersecurity and the social impact of emerging technology. LiveAIWire is an independent, human-led technology publication using AI-assisted research, editorial production and original AI-assisted editorial illustrations under his direction.