Regulating robo-advice for consumers’ financial decisions: The interplay between algorithm quality & digital choice architecture

Abstract

Rapid advances in digital interfaces, the quality of predictive algorithms, and the availability of high-speed internet are allowing firms to offer robo-advice: automated online guidance about which financial products are suitable for a consumer. At their best, robo-advisors feature algorithms that accurately predict which financial products are the optimal match for a given consumer’s needs and use digital choice architectures (that is, ways of presenting the product recommendations online) that are designed to strongly guide consumers to those best-fit products. Yet when algorithms lack great accuracy, strong guidance in favor of products incorrectly predicted to have the best fit could lead consumers to select options that do not meet their needs well. The possibility that strong guidance results in different financial outcomes depending on the quality of the algorithms has important implications for regulating robo-advice. We propose that, consistent with current regulatory practice for human advisors, regulators should require any firm wishing to implement a robo-advisor for its clients to demonstrate both that the robo-advisor is honest (that is, it works in the best interests of the client) and that its algorithms meet some established level of accuracy. In addition, regulators should require any firm offering a robo-advisor to demonstrate that the strength of the guidance provided by the digital choice architecture aligns with the predictive accuracy of the advisor’s algorithms. We offer several practical suggestions for implementing such a regulatory strategy. The need for regulations is becoming increasingly urgent because many robo-advisors now use artificial intelligence (AI) to make predictions. AI has made robo-advisors more powerful and easier to use, which is likely to expand their adoption—and, by extension, the harm that would be caused if subpar robo-advisors are allowed on the market.

Consumers often make costly mistakes when choosing financial products, such as health insurance or retirement investments. They err in part because these decisions are complex. Consumers may not fully understand the features of the products they are buying. Even when they do, they may be forced to make confusing trade-offs between features. In the case of health insurance, for example, people must decide whether paying higher premiums to have a lower deductible is worth the price. In the case of retirement investments, people need to decide whether the possibility of gaining the largest returns is worth the high risks that normally accompany investments that have the potential to generate such impressive profits. Mistakes can also occur because consumers typically make major financial decisions only rarely. Hence, their opportunities to learn from experience are limited.

Traditionally, many consumers seeking financial advice have relied on human advisors they think can help them make better financial decisions than they could make on their own. Recently, digital programs known as robo-advisors have become available to help with these decisions. A robo-advisor can be defined as any automated online service that uses predictive models to provide personalized guidance for purchasing financial products.^1,2 Today, many robo-advisors use artificial intelligence (AI) to make those predictions; that is, at least some part of their prediction algorithms, or programs, learn from experience and draw conclusions without using rules spelled out by human programmers. Thanks to rapid advances in the quality of the algorithms used by robo-advisors and the availability of high-speed internet, robo-advisors can now predict with greater accuracy than a decade ago which financial products best match a given consumer’s needs and can present the results online in ways that smoothly guide the consumer to those products.^{2 –5}

For example, the robo-advisor Alex, sold to employers by Jellyvision, provides personalized health care plan recommendations to employees, basing the advice on an individual’s own input (such as the individual’s projected health care needs) and on data relating to the use of medical services and the ultimate health care expenditures of large numbers of people. ⁶ In the investment sector, Vanguard’s Digital Advisor provides personalized advice on investment choices and allocations, taking into account an individual’s preferences (such as investment goals, risk tolerance, and stage in life) as well as large-scale data on the past performance of the company’s investment options. ⁷

In simple terms, robo-advisors’ algorithms can be thought of as prediction machines that produce a forecast of how well different financial products fit a given consumer’s needs. ⁸ On the basis of this predicted fit, robo-advisors will recommend one or more financial products. Ideally, the predicted fit should align perfectly with the consumer’s best interests. In reality, that ideal may not be fully achievable. Even so, if the algorithms are at least as accurate as skilled human advisors and if the robo-advisors are programmed to prioritize the consumer’s benefit, people who follow their advice can do as well financially—and at lower cost and with less effort—as individuals who rely on an honest, skilled human advisor. Thus, robo-advisors can potentially bring high-quality financial advice to many people who could not afford or would not consider consulting a human advisor and, in the process, can reduce financial inequities.

However, various factors could cause robo-advisors to lead consumers to products ill-matched to their needs. If the use of poorly performing robo-advisors became widespread, their failures would undermine the finances of many more people than the number who are now harmed by human advisors. To help avoid that outcome, we offer suggestions in this article for designing consumer protection regulations specific to robo-advisors. Our suggestions are based on an understanding of the ways high-quality robo-advisors should function and where they can fall short.

The escalating use of AI by robo-advisors increases the urgency of the need for regulation for two main reasons. For one, compared with traditional predictive analytics, the inner workings of AI programs are more difficult to understand, which makes foreseeing possible adverse outcomes more challenging for the government agencies responsible for protecting consumers. In addition, because AI makes robo-advisors more powerful and easy to use, more and more people are likely to flock to them, which means that ever larger numbers of people could be harmed by any failures or weaknesses.

How Robo-Advisors Affect Outcomes for Consumers

How well a robo-advisor meets an individual’s financial needs depends in large part on three factors: (a) the quality of the algorithm for matching individual preferences to products, (b) the robo-advisor’s choice architecture (the way the online interface presents product recommendations to the consumer), and (c) the interplay between the algorithm and the choice architecture.

One way to evaluate the quality of an algorithm is to look at how closely its predicted fit between a product and a given consumer corresponds to the fit that would be predicted by honest, independent human experts who had full information about the person and the product—that is, who knew the consumer’s situation and preferences; the relevant financial products available in the market; and the projections of relevant future developments for the consumer, the market, and the market environment. The quality of algorithms can be considered high if the recommendations of the algorithms and those of the expert human advisors closely correspond to each other.

An algorithm could miss that high mark because of bias or noise.^9,10 Predictions would be biased if an algorithm systematically pointed consumers to products that were not in the consumers’ best interests. Biased predictions might arise, for example, from biases in the data set on which the algorithm is based or from biases in the goals programmed into the algorithms. The use of biased algorithms can occur unintentionally, such as when the creators of a robo-advisor are unaware of biases in the data set used when developing its algorithm. But biased algorithms can also be applied intentionally. After all, as is true of traditional human-based financial advice, robo-advice is sometimes designed to manipulate the consumer.^1,11 Companies develop algorithms—and digital choice architectures—for their own purposes, and some firms may want to maximize their immediate profits rather than promote long-term consumer interests. Some firms, with the goal of cost cutting, may economize on research and the quality of the algorithms they develop.

The predictions made by algorithms would be considered noisy if they deviated from the consumers’ best interests randomly. Such noise can arise when an algorithm does not capture some predictive variables that affect product fit. Although algorithms can potentially be built to reduce both bias and noise, distortions are nonetheless likely to occur unintentionally and could lead to suboptimal financial advice.

As for the influence of choice architecture, the ways product options are presented can affect which one a consumer chooses. For instance, positioning an option at the top of a page above other choices will tend to sway decisions in favor of that top-positioned choice. The ideal choice architecture helps consumers make the best decisions for themselves with the least effort. ¹²

Choice architectures for robo-advisors can be designed to adjust how powerfully they steer someone toward an algorithm’s best-fit options.^{12 –14} Guidance that is strong makes it highly likely that a particular product will be chosen. For instance, the interface might set a single product as a default. Default choice architectures often have a relatively strong impact on the choices consumers make, ¹⁵ mainly because people like to take the path of least resistance and have to make an effort to find alternatives when they do not accept a given default, but also because consumers may assume that preset defaults are quality recommendations. The stronger the choice architecture guidance, the more effort a consumer would have to exert to deviate from the guidance.

Intermediately strong guidance is less directive. Here, the interface might present consumers with a set of options ordered in terms of predicted fit, ¹⁶ which enables consumers to easily inspect and compare multiple alternatives.

Weak guidance might offer little or no direction, such as by merely translating descriptions of complex products into easy-to-understand terminology, using simple diagrams instead of words to present information, and not presenting the products in a particular order. ¹⁷ This type of information-based guidance typically affects consumers’ choices relatively weakly. ¹⁴

Recent research we conducted shows that beyond the independent effects of algorithm quality and the strength of the choice architecture guidance, the interplay between these two aspects of robo-advice also influences consumers’ responses to the advice. ¹⁶ In particular, we observed that when algorithm quality is high, using a choice architecture that guides a person strongly to a specific option is helpful for consumers. Directing attention toward the predicted best-fitting financial products helps consumers avoid choosing a suboptimal product. However, when algorithm quality is low, using a choice architecture that delivers strong guidance may be harmful: Because the products recommended by the algorithm may not, in fact, match customer needs closely, consumers could end up making worse decisions than they would have made had they researched financial products themselves.

Goals for Regulating Robo-Advice

As we have noted, our major aim in this article is to spell out the regulatory implications of the ways the quality of robo-advisors’ algorithms, the strength of the guidance provided by their choice architectures, and the interplay of those factors can affect the financial health of consumers who take advice from robo-advisors. To that end, following are some goals we have identified for robo-advice regulation.

Some regulations for robo-advice can be modeled after the policies that regulators, such as the Securities and Exchange Commission in the United States, have already developed to prevent human financial advisors from harming customers. ¹⁸ These policies were adopted with the goal of ensuring that financial advice is honest, in that it is not systematically biased in favor of the advisor’s or a supplier’s profitability and places the consumer’s interest first. Regulatory policies for human advisors are aimed at ensuring that advice is not only honest but also sufficiently accurate. Poor training of a financial advisor, for instance, is no excuse for guiding consumers to low-quality products. However, regulators have recognized that financial advisors cannot be asked to guarantee that they recommend the single best product to a consumer, given that multiple products may serve the consumer’s best interests or that it may be impossible to know with certainty which product is the best option for a specific consumer. ¹⁸ For example, regulations regarding certain retirement plans prohibit advisors from recommending financial products that are objectively low performing or more costly than other equivalent products, but the regulations do not set requirements for or define what would be the best product to recommend to consumers. ¹⁹

For the algorithms that feed robo-advice, a similar set of regulatory requirements can be formulated. We propose that these algorithms should, first, be honest in that the predictions of consumer–product fit should not be biased in favor of profiting the firm that owns the advisor or a product supplier. Rather, the consumer’s benefit should be the priority. Second, when the algorithms are honest, their predictive quality—their ability to predict good fits—should be no worse than that of independent human advisors who act on behalf of the consumer. Algorithms, like human advisors, cannot be expected to be perfect. The algorithms of robo-advisors that are designed to generate outputs benefiting consumers can nonetheless vary in the accuracy of their predictions of fit, depending on the richness of the data available and the degree to which consumers, markets, and the performances of financial products are noisy, as they inherently are. In other words, algorithms will vary in the accuracy of their predictions of consumer–product fit, but their accuracy should be sufficiently high to be helpful to consumers.

Regarding digital choice architecture, regulations would bar robo-advisors from presenting choices in ways that promote products that are profitable for the firm that owns the robo-advisor or for its suppliers when those products are not also the best fit for consumers. In addition, regulations should take into account the interrelationship between choice architecture guidance and algorithm quality. That is, a choice architecture would gain approval only when a firm’s algorithm met the honesty requirement and if the predictions of this algorithm were presented in a way that was calibrated to the accuracy of the predictions. More specifically, beyond demanding honesty, regulations for choice architecture design should require the strength of the guidance to be weaker when the demonstrated predictive accuracy of the algorithm is relatively low than when the predictive accuracy is higher. For example, regulations might authorize presenting a single financial product to the consumer only when the algorithm’s predictive accuracy is very high.

In summary, we propose that regulators should require robo-advisors to demonstrate (a) that their algorithms are honest and sufficiently accurate and (b) that their choice architecture guidance also is honest and, importantly, aligns well with the algorithms’ predictive accuracy. Choice architecture can provide strong guidance when accuracy is high, and weak choice architecture guidance should be used when accuracy is merely sufficient. Regulators should consider whether they wish to explicitly require strong guidance in the case of high-quality algorithms (to prevent consumer mistakes) or if strong guidance would merely be allowed in that circumstance.

Potential Practical Regulatory Strategies

How might federal or state regulators go about incorporating these ideas into concrete requirements for robo-advice? Several strategies could parallel practical regulatory strategies already used in the regulation of human financial services.

At one extreme, firms could be required only to accurately describe the objectives and mechanisms of the robo-advisors’ algorithms and choice architectures, with regulators merely auditing the accuracy of firms’ disclosures. At the other extreme, firms could be required to obtain prior approval of their combinations of algorithms and choice architectures, with regulators setting standards for approval, testing robo-advisors in advance, and auditing robo-advisors for continued compliance with the standards.

Intermediate options include having regulators set standards for the combinations of algorithms and choice architectures and auditing for compliance without requiring prior approval or advance testing. Another intermediate option could be giving negatively affected consumers private rights of action—the right to bring a lawsuit against a firm deploying a robo-advisor that does not meet regulatory standards. ²⁰

As a first step, to set standards for algorithms’ honesty and accuracy, regulators could adopt an approach similar to the use of financial risk stress tests: They could define a large sample of hypothetical individuals and require a robo-advisor to provide recommendations that take into account numerous projections about world or market events. The regulators would then compare the resulting advice with that of an independent panel of experts who kept the consumer’s best interests in mind.^21,22 For example, regulators could judge the honesty and accuracy of a robo-advisor’s algorithm on the basis of the correlation between the robo-advisor’s predicted product fit and the fit that fully informed independent experts assigned to the same consumers. If the robo-advisor’s algorithm compared favorably, it would be approved. Potentially, the independent expert advice could also be used as a basis for developing a regulatory benchmark algorithm, thereby allowing this approach to be automated.

Next, regulators could develop standards for the combination of algorithms and choice architectures. One innovative idea for determining what the standards for robo-advisors would be is for regulators to sponsor contests in which robo-advisors are ranked according to their performance—for example, as judged by independent experts and consumers. Of course, the details would vary considerably by context, but rankings of health care providers and investments funds already exist and could serve as models.^23,24 The algorithm accuracy and choice architecture combination of the highest ranked robo-advisors could provide best practice examples to guide other robo-advisors and inform the development of choice architecture standards by regulators.

A second innovative idea is to assess the combination of algorithm and choice architecture post hoc, that is, after robo-advisors have been on the market. Regulators could require firms to maintain a record of all their robo-advice—similar to how flight recorders (colloquially known as “black boxes”) in airplanes record pilot voice and flight data. The firms would then make that record available to regulators, who would audit the record for the honesty and accuracy of the robo-advisor at the time the advice was given and would assess whether the choice architecture was appropriate given the algorithms’ accuracy. ¹ This approach would be similar in spirit to proposals for increasing the kinds of crash data that self-driving cars are required to store to enable investigators to assess the appropriateness of the self-driving system’s response to the traffic situation that ended in a collision. ²⁵

The content of any regulatory standards would necessarily vary by the financial service domain, with honesty as a universal goal and the specificity and strictness of competence standards depending on the maturity of the technology and the degree of regulators’ confidence in the robo-advisors’ judgments. Regardless of the regulatory approach, it is important to recognize that the same advanced technology, such as AI, and the availability of large data sets that make robo-advice such a potentially powerful tool for improving consumer welfare at scale can also benefit regulators, because they, too, can draw on advanced algorithms and big data to develop and enforce standards.

Conclusion

The interplay between algorithm quality and digital choice architecture has important implications for the financial outcomes of consumers who use robo-advisors and thereby also for the regulation of those advisors. In this article, we have drawn on our recent behavioral science research into the effects of strong and weak choice architecture guidance to offer several practical suggestions for regulatory strategies. ¹⁶ Many opportunities exist at the intersection of behavioral research and legal research for further empirical and legal study of how best to regulate robo-advice about financial products. Following are two such opportunities.

First, from the empirical perspective, the accuracy with which algorithms match consumers’ needs and available products and how effectively the digital choice architecture guides consumers to particular financial products are likely to differ depending on the individuals and the products. Predictions of the fit between an individual and the available products may be easier to make for some individuals than others, and some individuals may benefit more than others from choice architecture guidance.^26,27 Future research could address how regulations for robo-advice could take these differences into account at different stages of the technology’s development.

Second, in the legal domain, the need to use a digital choice architecture appropriate for the accuracy of the algorithm making predictions seems to suggest a potential new regulatory criterion for financial advice that may be appropriate for robo- and human advisors alike as they both increasingly rely on AI technology to support their advice. This criterion can perhaps be described in general terms as “appropriate or justified guiding force in providing advice.” Such a criterion would require that the firm offering advice demonstrate that its human communication or digital choice architecture is well aligned with the predictive accuracy of the underlying advice. It would be interesting to study how such a regulatory criterion can best be formulated and how effective its use will be in regulatory and advisory practice.

In this article, we have offered several practical suggestions for implementing a regulatory strategy for robo-advisors. We are optimistic that future researchers can test the behaviorally and empirically informed strategies that we and others have proposed and apply them in a variety of other financial product domains.

Suggestions for Policymakers, in Brief

• As is true for human financial advisors, robo-advisors should be required to demonstrate that they are honest (that is, they work in the best interests of the client) and that their algorithms meet a minimum level of accuracy.

• To help set standards for honesty and accuracy, regulators could adopt an approach similar to those of financial-risk stress tests. That is, they would compare robo-advice with that provided by an independent panel of experts who kept the consumer’s best interests in mind and would require a set level of agreement with the experts.

• Regulators should also require a demonstration that the strength of the guidance provided by the robo-advisor’s digital choice architecture aligns with the predictive accuracy of the robo-advisor’s algorithm.

• To set standards for such alignment, regulators could sponsor contests in which robo-advisors provide advice to the members of a consumer panel. This advice would be evaluated to see which combinations of algorithms and choice architectures were most beneficial for consumers. Regulators could then develop best practice guidance based on the findings.

• Akin to the way airlines are required to capture flight data with flight recorders, regulators could also require robo-advisors to record all their advice so that it could be evaluated if charges of dishonesty or incompetence were brought.