Can AI Replace Your Financial Adviser? Not Yet. But Wait.

Effective financial advice requires five financial qualities. Here’s how ChatGPT stacks up.
Shlomo Benartzi

Although it has been less than a year since ChatGPT was publicly released, the AI tool and its competitors have already achieved some impressive milestones. They’ve shown they can pass the bar exam for lawyers and help solve some of the hardest medical cases.

Is AI ready to replace your financial adviser?

The advantages are obvious. Professional financial advice is currently costly, and beyond the reach of many Americans. AI could drive those costs down and make smart, personalized guidance available for everyone 24/7.

AI can also expand the range of financial decisions covered by advisors. In the 21st century, people don’t just need help mixing ETFs into a portfolio—they also have to make hard choices about savings, insurance and debt management. Intelligent software can help in these situations, whether it’s choosing when to refinance a mortgage, or selecting the best health plan.

Of course, before we unleash AI on our finances, it’s essential that we understand the strength and weaknesses of these software tools. To do that, let’s look at five essential qualities for effective financial advice and see how AI currently stacks up. When these AI tools are lacking the requisite abilities, I’ll look at what it would take for AI to get there.

1. Debiasing

Let’s start with the bad news.

One of the primary value-adds of a financial adviser is debiasing, or helping clients avoid costly mistakes caused by behavioral tendencies. Consider myopic loss aversion, which causes people to overweigh short-term losses and invest too conservatively, even when their investment horizon is 30 years or longer. In one study I conducted with Richard Thaler, those who were shown a one-year chart of investment returns allocated 40% of their portfolio to stocks. In contrast, those who were shown long-term charts allocated 90% of their portfolio to stocks.

A good adviser can help people make financial decisions that align with their long-term goals. They steer clients away from those short-term charts, or the latest market swings that constantly pop up on mobile phones, and help clients choose investments that fit their actual time horizons.

Unfortunately, a working paper led by Yang Chen at Queens University in Canada showed that ChatGPT exhibits many of the same behavioral tendencies and biases that a good adviser tries to minimize.

For example, humans tend to choose riskier options after experiencing losses, as they try and break even. In Las Vegas, this is known as doubling down. ChatGPT suffers from the same tendency, which could lead to costly mistakes. If an investor lost a lot of money after the crypto crash, ChatGPT could recommend that they buy even more crypto, doubling down on the risky asset.

And it gets worse. That’s because AI tools are also highly overconfident. It’s not that they get it wrong sometimes—it’s that too often they think they’re right. This can amplify existing biases, as the software not only fails to self-correct, but can give its human clients a false sense of security.

To improve the performance of AI advisers, we need to create metarules—that’s a rule that governs other rules—to help the software override these biases. Metarules have already been successfully deployed to help AI models like ChatGPT minimize their racism.

We need to apply a similar strategy to these other behavioral biases, such as the deductibility bias and overconfidence. One possible approach is to have the AI, whenever it recommends a specific financial action, also review reasons why that action might be a mistake. It’s like an internal audit, forcing the software to consider what it might have missed.

To understand why metarules are often necessary, it’s important to understand how these AI tools learn. They are known as large language models, or LLMs, and they are trained on massive datasets of text pulled from the internet. Because the internet often represents human nature in an unfiltered form, the software reflects many of our lesser impulses and tendencies.

The good news is that AIs are almost certainly easier to debias than humans by applying metarules. We can’t directly edit the software running inside our head. But we can revise our AI models.

2. Empathy

The next key quality for an adviser is empathy. Consider, for example, an investor who’s nervous and anxious about market volatility. Research shows that the background mood of investors can have a powerful impact on their financial decisions with fear driving risk avoidance, and anger leading to more risk-taking. The role of a good adviser is to reassure and support during market turmoil, so that fear and other emotions won’t damage our long-term financial prospects.

The good news is that ChatGPT excels at empathy. One recent study compared the responses of ChatGPT and human doctors to the questions of real patients that had been posted on an online forum. The answers were then evaluated by a panel of healthcare professionals, both in terms of quality of information and empathy.

The results were a resounding win for AI. The healthcare professionals were nearly four times more likely to say that the ChatGPT responses provided “good or very good” information. But they were nearly 10 times more likely to say that ChatGPT was empathetic. Specifically, 45% of AI responses were rated as empathetic or very empathetic, compared to only 4.6% of physician responses.

These results suggest that there are some critical adviser tasks that AI can already perform extremely well; the future is here. Like busy doctors during a pandemic, many advisors don’t have the time or ability to reassure their clients during market corrections. The technology can help them become more human, or at least scale their humanity.

The next time there’s a major market drop, advisors don’t have to be limited to making a few calls to their wealthiest clients. Instead, AI can deliver empathetic responses tailored to each client. For instance, if a client is an experienced investor and checks their portfolio on a daily basis, the AI can provide reassuring data about long-term market trends, as well as the costly impact of market timing.  

3. Accuracy

Another important adviser quality is getting the facts right. Even if AI can be debiased, it still needs to base its advice on accurate representations about investments, inflation, taxes and more.

More bad news, at least for those dreaming of an AI adviser: The bots are currently very unreliable and make lots of mistakes. When I asked the tool to solve 12x23x34, it told me the answer is 9,336, which is close but wrong. I then gave it a second chance. It came up with a different wrong answer.

But AI isn’t just inaccurate—it often hallucinates and makes up facts. (Just ask the lawyer who was recently fined $5,000 for using ChatGPT for legal research, only to discover that the AI invented most of his citations.)

For instance, when I asked a leading AI tool to help me choose between Vanguard and Fidelity Nasdaq index funds, it came up with a very impressive answer focused on their long-term performance and expense ratios. I almost believed its confident recommendation that Vanguard was a better choice. The only problem was that it used the wrong funds as the basis for its analysis, using numbers from a Vanguard S&P 500 fund and a Fidelity real-estate fund. It was both highly confident and completely inaccurate.

This problem can be largely solved with plug-ins, or external tools that the AI calls upon to supplement its known weaknesses. When you ask Google a math question, it pulls up a calculator alongside the answer; AI tools should do the same thing. In addition to using a calculator, AI advisers should be integrated with reliable financial databases, such as Morningstar, that can ensure that its models and recommendations are based on accurate representations of the financial world.

“People too often think of language models as complete solutions to any problem, rather than as components in intelligent applications,” says Dan Goldstein, a senior principal researcher at Microsoft Research, specializing in AI and human-computer interaction. “The optimized systems and vast data stores of the financial world won’t be replaced by AI—they’ll be called upon by AI.”

4. Best interest

Advisers must act in the best interest of their clients. They can’t, for instance, recommend a more expensive fund class just because it makes them more money. In theory, then, AI should be less likely to get into conflicts of interest. Unlike humans, ChatGPT isn’t trying to maximize its income.  

But that’s just theory—we don’t really know how well AI will perform. One possibility is that it will have similar issues to humans. For instance, a study by Brad Barber, Terrance Odean and Lu Zheng found that investors are more likely to buy mutual funds with higher marketing expenses, even when those expenses reduce their overall performance through higher fees. (These funds are likely worse investments, but consumers are influenced by their advertising.) WHY SO? AI could fall into the same trap, as funds that spend more on marketing could loom larger in the AI database.

Given this uncertainty, it’s important that AI architects audit the recommendations of the digital adviser. This is similar to a meta-rule, just instead of erasing bias it’s focused on erasing conflicts of interest.

Fortunately, it’s likely that AI is easier to monitor for conflicts of interest than a human adviser. If the software starts recommending investments with high-fees or mortgages with high interest rates when there are cheaper alternatives, the AI tools might even be able to auto-correct, like spell-check fixing a typo.

Goldstein believes one key is emphasizing transparency, which is only possible with AI advisers. “When decisions are made behind closed doors, we can only wonder about some of these issues,” he says. “But when the inputs and outputs of every decision are logged, they can be put through checks that were never before possible.”

5. Consistency

Good financial advice should be consistent. That is, if the same client takes the same portfolio to different advisers, they should offer similar advice, focused on the same time-tested principles.

While ChatGPT gave me different wrong answers to the same basic math question above, this should be a fixable problem. AI advice should be able to achieve consistency by confirming that it gave the same advice to clients with similar financial needs and preferences.

Human advisors, however, could use help, as research suggests they struggle to offer advice that consistently reflects the goals, circumstances and preferences of their clients.

One recent study, by Juhani Linnainmaa, Brian Melzer and Alessandro Previtero, showed that clients tend to invest in funds with different fees and risk profiles after their adviser dies or retires, and they are placed with a new, randomly selected, adviser. This isn’t because their investment preferences suddenly changed—it’s because their new adviser inflicted his or her own beliefs on their portfolios. If the new adviser selected risky investments for his own personal portfolio, or expensive funds, he assumed his clients would prefer that too.

Once AI tools achieve consistency, they shouldn’t exhibit this problem. The software should deliver the same advice to clients in the same situation, much as Netflix recommends similar content to people with the same viewing history.  

What the future could look like

A lot of improvements are needed before AI can become an effective financial adviser. Nevertheless, it’s clear that AI will play an important role in the future of financial advice.

What might this future look like?

One potential model comes from the medical domain, where smart software and doctors have been working together for years as a hybrid team. In particular, doctors increasingly rely on clinical decision-systems to help them improve their quality of care, as the digital tools can reduce misdiagnoses or shorten the time to make a diagnosis.

New research extends this model to AI. One study by doctors at Beth Israel Deaconess Medical Center in Boston found that ChatGPT could solve nearly 40% of the hardest medical cases, which is a higher success rate than most human doctors working alone. Even more impressive, AI included the correct diagnosis on its longer list of possible diagnoses 64% of the time.

Of course, a human doctor is still required to filter the extended list generated by ChatGPT, and select the best diagnosis. This suggests that the AI can help us expand our thinking, even when it can’t actually find the answer by itself.

While there are no studies of the quality of hybrid financial advice, I speculate that the hybrid model will win, provided humans learn how to effectively collaborate with AI. But even if I’m wrong and pure AI ends up providing the best quality of advice, there’s a behavioral reason why clients are still likely to prefer a hybrid approach.

Consider the self-driving car—it’s not enough for the auto-pilot software to drive far better than human drivers. It has to be nearly perfect, or else it’s likely to be rejected by users. This is known as algorithm aversion, and applies to algorithms in many domains, including finance. Algorithm aversion suggests that we’re not mentally ready for pure AI advisers. We’ll only take financial advice from AI that is monitored by other people, much as people expect a pilot to oversee the auto-pilot in the cockpit.

While a hybrid approach is likely to improve the quality of financial advice, it can also dramatically increase access to advice. Currently, an industry survey by MagnifyMoney found that only about 35% of men and 25% of women have access to financial adviser. My hope is that human advisers use AI to help them serve more people, so they can help at least 50% of Americans improve their financial wellbeing.

And what about those Americans who still won’t be able afford a human adviser? I believe AI can be used to deliver advice 24/7, provided we fix those critical issues involving accuracy and debiasing. To help people deal with algorithm aversion, we can offer human supervision as needed.

If you’re a financial adviser, I wouldn’t worry about losing your job to ChatGPT. (Auto-pilots didn’t put pilots out of work.) Instead, I’d focus on how you can use the technology to deliver better advice to even more people.

Americans need effective, objective and empathetic advice. AI is far from perfect, but with a few important fixes it can become a critical tool for helping those who need it most.

The Wall Street Journal