Does your robo-advisor know you’re NOT a 30-year-old millionaire?

Artificial intelligence can successfully understand how different people think about investing—accurately capturing preferences across gender, age, and income, according to a new UC Berkeley Haas study. But without explicit prompting, AI defaults to mimicking young, wealthy men.  

The working paper, co-authored by professors Anastassia Fedyk, Ali Kakhbod, and Ulrike Malmendier, and PhD student Peiyao Li, points to both the promise and peril of automated investment advice platforms dubbed “robo-advisors.” These automated advisors have the potential to democratize financial guidance, but when left to default settings, deeply embedded algorithmic biases are likely to reinforce the very inequalities that have historically excluded women, older adults, and lower-income investors.

“We have no doubt the industry can design systems capable of responding to a growing and increasingly diverse pool of retail investors,” said Malmendier, the Cora Jane Flood Professor of Finance at Haas and faculty director of the O’Donnell Center for Behavioral Finance. “But these systems must be intentionally designed to serve all investors fairly.”

The findings provide reason to be optimistic about the future of automated investment advice platforms, which have grown more than tenfold over the past decade. These AI-powered services deliver algorithm-driven financial advice at a fraction of traditional advisors’ costs. 

“A big promise is that generative AI can do more than offer one-size-fits-all advice—when it’s prompted to reflect who it’s speaking to, it can mirror the reasoning people give for their views and help distinguish gaps in understanding from differences in preferences,” said Kakhbod, an assistant professor of finance at Haas.

But the research arrives amid growing concerns about algorithmic bias—the tendency of computer code to perform better for some demographic groups than others. Medical device software has been shown to disadvantage certain demographic groups, facial recognition software mistakes people of color more frequently, and lending algorithms have given white borrowers better interest rates than equally qualified minorities.

Answering the key question

To explore whether robo-advisors might one day provide sound advice to investors with a wide variety of needs, goals, and perspectives, the researchers had to first answer a key question: Could AI truly understand the diverse preferences of actual investors?  

To find out, they ran a carefully designed experiment. First, they surveyed more than 1,200 people mirroring U.S. demographics, asking them to rate stocks, bonds, and cash as investment options. The researchers then asked Open AI’s GPT-4 to role-play as survey participants with different demographic characteristics. 

The results were surprisingly accurate. GPT-4 correctly captured how men rate stocks higher than women do, how higher-income individuals prefer stocks and bonds over cash and that older investors favor cash holdings. When the researchers compared average ratings across eight demographic groups (split by gender, age, and income), they found substantial agreement between human and AI responses, with correlation coefficients ranging from 0.57 to 0.78.

What’s more, AI didn’t just mimic the ratings; it captured the thinking behind them. By scrutinizing the free-form explanations using sophisticated text analysis, the research team found that both humans and GPT-4 discussed and agreed upon the same core concepts—risk, return, knowledge, and experience—with AI responses essentially matching human judgments.

“AI didn’t just mimic numerical ratings—it replicated people’s reasoning, which is really promising for taking that next step in constructing AI advice platforms,” said Fedyk, an assistant professor of finance at Haas.

The model also correctly captured the interplay between different factors driving investment attitudes. The degree to which market knowledge and positive experiences correlated with higher return expectations and lower risk perception was nearly identical in both human and AI responses.

A troubling bias 

But the research uncovered a troubling finding. When the researchers didn’t specify demographics in their AI prompts—simply asking GPT-4 to role-play as “an online survey participant” without mentioning gender, age, or income—the responses became heavily skewed.

Without demographic guidance, 55% of AI responses resembled those of young, high-income men, even though this group represents less than 15% of the population. All AI responses reported incomes above $54,000, and only 6% identified as being 39 years old or older.

This “default setting” can have real implications for retail investors. If robo-advisors are not prompted with specific user demographics, they risk potentially steering women, older adults, and middle-income investors toward unsuitable strategies.

“This is a major risk, as it creates systematic bias precisely where robo-advice is supposed to broaden access,” Kakhbod said.

“It can potentially lead somebody into an investment that has a level of risk beyond their tolerance. And that could discourage them from investing in the future,” Fedyk said.

An unexpected advantage

On the other hand, researchers uncovered one unexpected advantage of AI: GPT-4 was more logically consistent than humans. Survey participants ranked pairs of investments: stocks versus bonds, stocks versus cash, and bonds versus cash. Economic theory says these rankings should be “transitive”—if you prefer stocks to bonds and bonds to cash, you should prefer stocks to cash.

“AI’s consistency—its choices are almost always logically transitive—can help us identify when the binding constraint is understanding, not preferences,” Kakhbod said.

GPT-4’s rankings were transitive 98.7% of the time. Humans? Just 84.4%. The gap was particularly pronounced among women, who more often indicated indifference between options and then violated transitivity in their subsequent comparisons.

“Humans are not always logically consistent, and we found this was largely driven by a lack of financial knowledge. The models are more consistent. That means AI could potentially help investors make better decisions,” Fedyk said.