☀ Key Stats
◆ Prompts written by people without previous AI-finance experience generated recommendations that produced 5.7% less simulated wealth at age 60 than prompts written by previous users.
◆ The difference was approximately $99,800, measured in 2025 dollars.
◆ Prompts written by women produced recommendations associated with 4.1% less simulated wealth at age 60, or approximately $59,900, than prompts written by men.
◆ Prompts from people with lower financial-literacy scores produced 4.1% less simulated wealth, or approximately $45,900, than prompts from the higher-literacy group.
◆ Women’s prompts generated an average recommended equity allocation that was 2.94 percentage points lowerbetween ages 22 and 60.
◆ Approximately two-thirds of the gender difference in diversified-equity advice came from differences in what men and women wrote in their prompts.
◆ The remaining one-third came from the model giving different recommendations to otherwise identical prompts labelled as belonging to a man or a woman.
◆ The researchers began with 1,000 U.S. survey responses, of which 952 passed authenticity and minimum-specificity checks.
◆ Approximately 48% of respondents said they had used an AI tool for financial advice or information during the previous three months.
◆ One-shot AI advice increased simulated stock-market participation by approximately 5–10 percentage points and average equity allocations by about 5 percentage points compared with respondents’ reported behavior.
◆ In the full life-cycle simulation, virtually all simulated individuals accumulated more than $10,000 in financial wealth by age 30.
◆ About 31% of recommended saving rates were multiples of 10%, while more than 98% of retirement withdrawals were no greater than 4% of wealth.
Note: These are simulated outcomes based on the assumption that people follow the AI’s recommendations throughout their lives. The researchers did not observe respondents making these choices or accumulating the reported wealth.
Continue reading ↓
The wording people used to request AI financial advice produced simulated wealth differences of as much as 5.7% by age 60, according to a new research paper.
The largest gap appeared between prompts written by people with and without previous experience using AI for financial information or advice.
Recommendations based on prompts from people without previous AI experience produced approximately $99,800 less simulated wealth at age 60.
That was 5.71% below the result generated by prompts from respondents who had previously used AI for financial guidance.
The results do not show that using AI increases a person’s wealth.
They show how differences in the questions people wrote—and differences in how the model answered them—accumulated when the researchers simulated following the advice throughout adulthood.
Women’s prompts produced 4.1% less simulated wealth
Prompts written by women generated recommendations that produced approximately $59,900 less wealth at age 60 than prompts written by men.
That represented a difference of 4.1%.
The gap came mainly from investment allocations rather than saving.
The model recommended an average equity share between ages 22 and 60 that was 2.94 percentage points lower when advice was generated from women’s prompts.
The future value of cumulative saving was nearly the same for the two groups.
The researchers also found that the equity-allocation difference appeared across age and income categories.
Both the prompts and gender labels changed the advice
The researchers designed a separate experiment to distinguish between differences in what respondents wrote and differences in how the model responded to demographic information.
They selected the 83% of prompts that did not directly or indirectly disclose the writer’s gender.
The researchers then randomly added either “I am a man” or “I am a woman” to the beginning of each prompt.
Moving from a prompt written by and labelled as male to one written by and labelled as female reduced the recommended diversified-equity allocation by 1.50 percentage points in that period.
About 0.96 percentage points of that difference came from the content of the original prompts.
The remaining 0.54 percentage points came from changing the gender label while holding the rest of the prompt constant.
The study therefore attributes approximately two-thirds of the combined difference to what men and women wrote and one-third to the model’s response to the gender label.
Women more frequently wrote about subjects including family expenses, groceries, loans, credit and debt.
Men more frequently used terms related to portfolios, investment strategies, equities and cryptocurrency.
Those differences may convey relevant financial circumstances, but the experiment also found that the model changed its advice when only the gender label changed.
Financial literacy produced another 4.1% gap
Prompts from respondents with lower financial-literacy scores produced approximately $45,900 less simulated wealth at age 60.
That represented another 4.1% difference.
The researchers classified respondents as having lower financial literacy when they answered at least one question incorrectly on a five-question test.
The simulated wealth gap was driven mainly by portfolio allocation.
Prompts from the lower-literacy group produced an average equity allocation that was 1.64 percentage points lowerbetween ages 22 and 60.
The lower-literacy group was advised to save slightly more, which partly offset the effect of its lower equity allocation.
The study does not show that financial knowledge itself caused the simulated gap.
Differences in what the groups asked, the information they supplied and the model’s responses could all contribute to the result.
Previous AI users wrote prompts that generated more saving
Approximately 48% of respondents said they had used an AI tool for financial advice or information during the previous three months.
Prompts from those respondents produced substantially more simulated saving than prompts from non-users.
The future value of cumulative net saving through age 60 was 4.32% lower in the simulation using prompts from people without previous AI-finance experience.
The difference in equity allocation between previous users and non-users was small.
That means the 5.7% wealth gap was driven mainly by saving rather than investment allocation.
Previous AI use may have helped respondents write more detailed or effective questions.
It may also reflect other differences between the groups that the study was not designed to interpret as causal.
AI advice increased diversified investing
The average survey respondent reported holding about 30% of financial wealth in equities, while approximately one-third held no equities.
Applying the AI advice once increased simulated stock-market participation by about 5–10 percentage points.
It also increased the average equity share by approximately 5 percentage points.
When the researchers simulated following AI advice every year, stock-market participation became nearly universal.
Equity allocations increased by as much as 40 percentage points relative to respondents’ reported behavior in some age groups.
The recommended equity share averaged approximately 65%.
It rose during early working life and began declining after about age 45, although the decline was less pronounced than in a standard life-cycle investment model.
Most of the risky investment allocation went to diversified equity funds rather than individual stocks, cryptocurrency, commodities or collectibles.
The advice emphasized emergency savings
More than 20% of respondents in every age group reported holding less than $10,000 in financial wealth.
In the full simulation, virtually all individuals built a financial buffer above that level by age 30.
The model frequently introduced liquidity and emergency savings even when respondents did not mention them.
Liquidity appeared in 84% of the AI responses, compared with only 6% of respondents’ prompts.
High-yield savings accounts appeared in 59% of responses, compared with 3.5% of prompts.
Bonds appeared in 55% of responses, while Treasury securities appeared in 40%.
The researchers also found that the model sometimes named specific financial-product providers even when respondents did not request a branded product.
The recommendations often relied on simple rules
Although the advice moved simulated behavior closer to several broad principles of life-cycle finance, it did not closely reproduce the researchers’ economic model.
About 31% of recommended saving rates were multiples of 10%.
Approximately 34% of saving amounts were multiples of $5,000.
More than 98% of the model’s recommended retirement withdrawal rates were no greater than 4% of assets.
That concentration resembles the commonly used 4% retirement-withdrawal rule rather than a recommendation recalculated for every person’s circumstances.
The advice also tended to allow portfolios to change passively with investment returns.
It focused more on where to place new savings than on actively rebalancing existing investments.
The researchers found that structured, academic-style prompts reduced some of the reliance on round-number saving and withdrawal rules.
They did not meaningfully reduce the model’s tendency to allow portfolios to drift.
The researchers simulated decisions from age 22 to 90
The study began with 1,000 complete responses from a demographically balanced U.S. sample recruited through the Prolific online platform.
After authenticity and specificity checks, 952 responses remained.
Each respondent wrote three prompts.
One described the person’s financial situation, another asked about spending or saving, and the third requested investment advice.
The researchers then placed the prompts into 12 groups based on employment status, age and income.
Each year, simulated individuals were matched with a prompt from a respondent with similar characteristics.
References to age, income, wealth and investments were replaced with the simulated person’s current information.
GPT-5.2 generated written financial guidance, which was then converted into numerical choices for spending, saving and asset allocation.
The researchers simulated the same 1,000 virtual individuals from age 22 until death at or before age 90.
The model included employment changes, unemployment, income shocks, investment returns, taxes, retirement at age 65 and mortality.
The main findings were quantitatively similar when the researchers used Gemini 3 Flash and GPT-5.6 Terra.
✦ Why it matters ✦
AI systems can provide financial information at a far lower cost than many traditional advisory services.
That accessibility could make financial guidance available to people who do not use a professional adviser.
The study suggests that the models often promote broad practices associated with conventional long-term planning, including diversified investing and emergency savings.
It also shows that access to the same AI system does not necessarily produce the same recommendations.
The wording, detail and subjects included in a prompt can influence what the model recommends.
Some of those differences may reflect legitimate variations in income, debt, family responsibilities or risk.
Others may arise because users with greater financial or AI experience know how to provide the information needed for more tailored answers.
The gender-label experiment raises a separate issue.
The model reduced its recommended diversified-equity allocation when an otherwise identical prompt was labelled as coming from a woman.
That finding does not establish how every AI system will behave.
It shows that models can use demographic information in ways that change financially important recommendations.
The study also highlights potential conflicts when an AI system names particular providers, funds or financial products.
The paper does not examine whether those recommendations are influenced by commercial arrangements or how users interpret branded suggestions.
ⓘ How to read the findings
The wealth figures are outputs from a life-cycle simulation, not actual changes observed in respondents’ bank or investment accounts.
The study estimates the consequences of following the recommendations exactly.
It does not measure whether people understand the advice, act on it, partially follow it or ignore it.
The results depend on assumptions about earnings, employment, asset returns, taxes, retirement and mortality.
Different assumptions could produce different wealth levels or gaps.
The sample broadly matched the U.S. population by age, gender and income, according to the paper.
It overrepresented unemployed respondents and underrepresented the lowest-income group and people older than 70.
Survey participants were recruited through an online platform, which may not represent everyone seeking financial guidance.
The study also converted written recommendations into numerical spending and investment choices.
That translation introduced some variation, although the researchers found that most variation came from generating new written advice rather than converting the text into numbers.
The paper is a working paper and should be treated as preliminary.
Its results should not be interpreted as personal investment guidance or evidence that following AI recommendations will increase wealth.
