FinRegLab Responds to Treasury Request for Information on Artificial Intelligence.

Department of the Treasury

1500 Pennsylvania Avenue, NW

Washington, D.C. 20220

Download PDF

Thank you for the opportunity to provide input on the uses of AI in the financial services sector and the opportunities and risks presented by AI developments and applications. We commend the Treasury Department’s ongoing engagement on this important and rapidly evolving topic.

FinRegLab is a nonprofit, nonpartisan innovation center that tests new technologies and data to inform public policy and drive the financial sector toward a responsible and inclusive financial marketplace. With our research insights, we facilitate discourse across the financial ecosystem to inform public policy and market practices. Financial inclusion is central to FinRegLab’s mission because an inclusive financial system helps to increase broader economic participation, financial health, and wealth building, particularly among historically marginalized and disadvantaged populations.

Much of our research has focused on the financial inclusion and fairness implications of data and machine learning (ML) in consumer and small business lending, where we have written in depth about potential benefits and risks, current practices, current risk mitigation strategies and tools, and potential market and policy evolution. We incorporate those reports by reference here,¹ in addition to highlighting a number of other AI use cases that have potentially important implications for financial inclusion. Our comment is split into three sections:

• Background on AI adoption and evolution in financial services
• AI use cases with particular implications for financial inclusion
• Next steps to facilitate adjusting market practices and regulatory frameworks.

Overall, we emphasize the importance of considering the impacts of AI adoption on historically underserved and under-resourced populations. To the extent that AI and more representative data sources can help to produce more individually tailored financial services and to substantially reduce the costs of delivery, these populations could derive substantial benefits from such innovations since their needs are often not well met by existing systems. However, market incentives may not lead to such use cases being prioritized relative to more profitable customer segments. In addition, the risk of unintended consequences can be particularly high for such populations to the extent that AI adoption increases fraud and identity theft, perpetuates and exacerbates historical disparities, or creates new barriers to inclusion. We urge the Treasury Department, other policymakers, and financial services stakeholders more broadly to monitor and prioritize initiatives to ensure that AI adoption improves access to responsible financial services over time.

Background: AI adoption and evolution in financial services

Artificial intelligence is a term coined in 1956 to describe computers that perform processes or tasks that “traditionally have required human intelligence.” Machine learning is often used to refer to the subset of artificial intelligence that gives “computers the ability to learn without being explicitly programmed.”² In practice, these terms are often used to describe a broad range of models and applications that may be applied to a variety of data sources (e.g., tabular financial data, text, images, and voice) for a wide range of purposes (including predictive analytics, web and information searches, autofill functions, and generating new content in response to queries). The level of human involvement also varies depending on the particular application and technique. At a minimum human oversight is critical in determining what data the algorithms are exposed to, choosing among different techniques and model architectures at the start, and validating, monitoring, and managing the models depending on the desired goals and the risks to address.

It is important to emphasize that ML/AI is not monolithic and that terminology can vary in different contexts (e.g., investor pitches compared to data science convenings compared to public policy debates), which can sometimes make it difficult to understand how much new innovations differ from older forms of statistical analysis and automation. The proposed definition in the Request for Information also poses this challenge, since it appears to include a wide range of automated decisioning.³ There also can be tremendous differences between sectors, use cases, and individual companies as to how ML/AI is deployed and managed.

The financial services sector has been an early adopter of statistical analysis and automation, including the use of machine learning models decades ago particularly in contexts such as fraud detection, cybersecurity, and market trading.⁴ Improvements in computing power, access to digital information sources, and data science techniques for diagnosing and managing models have encouraged adoption in additional use cases over time, including marketing, credit and insurance underwriting, servicing, and customer support functions.⁵ In recent years, use of machine learning models in credit underwriting has been a particularly significant development, both because of the potential to improve access to credit for millions of people who are difficult to underwrite using traditional techniques and data sources and because it has required lenders to grapple with a range of regulatory requirements including individualized consumer disclosures and fair lending laws in addition to more general risk management expectations. Given the consequences to both lenders and borrowers, use in underwriting has triggered substantially more scrutiny and debate than older applications such as fraud.⁶

In the past 18 months, interest in and debate about AI applications has accelerated in financial services just as it has in other sectors after the release of ChatGPT in November 2022 underscored the potential for large language models (LLMs) and other forms of generative AI to create new content (including text and images) by relating prompts to learned patterns in training data.⁷ The largest and most technologically savvy financial services providers are testing these techniques across a broad range of use cases, and adoption has begun for selective functions in areas such as fraud detection and back office operations. However, as noted in the Treasury Department’s recent cybersecurity report, financial services providers have tended to be substantially more cautious in adopting these models than other sectors, due to several considerations:

• Data management and hygiene: Large language models are typically trained on massive datasets scraped from large portions of the internet.⁸ They are requiring more training data and becoming more complex as models become more advanced. For example, GPT (2018) has 117 million parameters and was trained on five gigabytes of data, GPT-2 (2019) has 1.5 billion parameters and was trained on 40 gigabytes of data, and GPT-3 (2020) has 175 billion parameters and was trained on 45 terabytes of data.⁹ Particularly with regard to proprietary LLMs, this raises substantial concerns for financial services providers about intellectual property considerations and risks in exposing their own proprietary information to such models, as well as accuracy, bias, and explainability as discussed below.
• Accuracy, consistency, and bias: The broad-based nature of LLMs’ training data, the frequency of inaccurate answers (often called “hallucinations”), variations in output wording, and the risk that the models will perpetuate or even exacerbate bias are shaping financial services providers’ decisions about when and how to deploy these technologies. Where queries are inputted directly by consumers, those inquiries could also become a potential source of inconsistency and bias.¹⁰ Risks can be reduced by carefully curating and filtering training data, engaging in extensive fine-tuning processes, or the use of different embedded models to overcome hallucinations, but these measures can be expensive and their effectiveness is still being evaluated.¹¹
• Explainability: The size and complexity of LLMs makes it extremely difficult for data scientists to understand how the models are generating their answers for purposes of risk management, regulatory supervision, or consumer disclosure. While the broader technology community is working to develop techniques and strategies to address explainability concerns in the LLM context,¹² financial services providers are particularly sensitive to the lack of agreed-upon practices given that they are subject to a range of explainability expectations in different activities, particularly in credit-related applications.¹³

All of these factors add to the cost and risks of deployment, particularly with regard to direct-to-customer applications. While a variety of large financial services providers have begun testing versions of these technologies, many banks have banned the use of ChatGPT, Dolly, GitHub Copilot, and similar proprietary products due to risk and security concerns.¹⁴ Instead, many providers are working to use open source LLMs such as LLaMA, Mistral, and open source GPTs, which they often bring within their firewalls before conducting supplemental training with their own data to build products or use them for various applications such as text summarization, extraction, generation, prohibited language detection, sentiment analysis, and computer vision.¹⁵ Financial service providers are also designing many of their early deployments as “employee assistant” models that are designed to assist coders, customer services representatives, and other knowledge-based employees, without replacing those employees or interacting directly with customers.¹⁶

Financial institutions and other firms are also tailoring their specific model choices to manage risks for the particular task at hand. For example, where a task involves text generation, companies may select different models based on the size of the parameters. While a smaller model may not perform best in a general setting, it might excel in a specific use case and present less risk of hallucinations where it is trained on smaller but more curated data sources.¹⁷ Additionally, companies can use RAG (Retrieval-Augmented Generation) or fine-tune LLMs to reduce hallucinations, though these processes increase costs.¹⁸ For tasks such as text extraction and sentiment analysis, where hallucinations usually are not a concern but accuracy still matters, LLMs that have been pre-trained on large amounts of data can provide higher accuracy than traditional NLP models, but have downsides with regard to computational cost and complexity of deployment. If analyzing the broader context in which a word or phrase appears is important, smaller language models like BERT can be used in some contexts.¹⁹ Where data labels are available and context is not critical, convolutional neural networks (CNNs), recurrent neural networks (RNNs), and some other simpler, less expensive models can also be used for a range of NLP tasks.²⁰

AI use cases with significant inclusion implications

Although there are serious reliability issues to be sorted out with various forms of generative AI, the technology is causing stakeholders to consider potential advancements in various use cases that previously seemed many years away. We highlight three financial services use cases for evolving forms of AI with particularly important implications for financial inclusion of historically underserved and under-resourced populations.

A. Delivery of financial advice and coaching

Even prior to recent advancements in generative AI, the development and deployment of chatbots has been accelerating rapidly across various industries, particularly as customer service or sales agents. These chatbots can generally be categorized as either rule-based or generative AI-based.²¹ Rule-based chatbots may use either simple keywords, or a natural language processing engine, to determine what the user is asking (i.e. intent), then direct a user to a set of pre-written template responses. Generative-AI chatbots use deep learning architectures, particularly transformers, to generate responses. They generate responses dynamically based on the context and the responses are not pre-scripted. Some chatbot systems combine these approaches, for instance by pointing the user to pre-written content as well as generating commentary or a summary of the static content.

Like in other industries, customer service chatbots have become widespread in the financial industry, though these are almost entirely rule-based systems. All of the top 10 commercial banks have released such a chatbot, and about 37% of the US population is estimated to have interacted with a bank through a chatbot in 2022.²² Some banks have developed their own rule-based chatbots using first-party data such as customer conversation logs.²³ Models such as Capital One’s Eno, released in 2017, and Bank of America’s Erica, released in 2018, have grown in adoption since their launch. Bank of America reported that Erica took four years to reach 1 billion interactions, yet surpassed 2 billion just 18 months later.²⁴ Erica can perform a range of tasks, such as informing checking account-holders of their monthly subscription charges and of any bills due.

Other banks have partnered with third parties for these services. For example, Kasisto, a company that specializes in creating conversational financial assistants, was founded in 2013 and released its first model in 2018.²⁵ Kasisto has now partnered to deploy its system, Kasisto AI (KAI), with 49 financial institutions, including over 30 credit unions and community banks, and serves more than 50 million customers. In 2018, the company partnered with JP Morgan to release a KAI-powered assistant to help corporate clients in its treasury services division.²⁶

FinRegLab has spoken to some financial institutions that are testing the use of LLMs to process inquiries, though they continue to use rule-based outputs. In general, banks remain extremely averse to deploying fully LLM-generated responses in customer-facing external applications due to concerns about hallucinations, inconsistency in answer wording, and the potential for legal liability vis a vis consumers or regulators, in addition to reputational damage.²⁷

One fintech startup, Cleo, has attempted to resolve some of these concerns with a combination of rule-based and LLM-generated responses. Cleo seeks to reach a younger audience that may find financial advisors inaccessible or intimidating, by adopting what they refer to as a “big sister persona”,²⁸ offering features including “Roast Mode”, which makes fun of a user’s spending habits, and “Hype Mode,” which congratulates the user on their good financial decisions.²⁹ Like the above applications, the chatbot uses LLMs for understanding user intent. It then classifies the request: some queries (such as specific requests for financial advice) are given a pre-written response. Cleo advertises that they have multiple comedians on staff helping to write many of their responses. However, other responses receive a response generated with an pre-trained LLM model, such as GPT-4 or Claude, summarizing material from Cleo’s customer service knowledge base.³⁰

As financial chatbots become increasingly sophisticated, questions about how they present to consumers and how consumers react to their use raise important implications for issues of privacy, equity and financial inclusion. One study found that consumers’ trust in a brand decreases more when disclosing personal information to an AI compared to a human, because consumers infer that data shared with an AI system is more likely to be shared with a wider audience.³¹ This effect can be mitigated when a bot is more anthropomorphic (i.e. with more human-like dialogue, names, or avatars), or when a customer is assured that their confidentiality will be protected.

In some sensitive contexts, the non-human nature of a chatbot may be an asset. A study within the UK’s National Health Service system suggested that using personalized chatbots increased self-referrals for mental healthcare services, an effect that was especially pronounced among non-binary individuals.³² Research in a similar vein suggests that customers who are shopping for sensitive personal items, such as anti-diarrheal pills or condoms, prefer to communicate with a chatbot over a human, while also favoring a less anthropomorphic chatbot that is clearly identifiable as such over a more human-like bot.³³

This research suggests that chatbots may be especially well-received for assisting customers in contexts where self-presentation concerns are high. Counseling about financial distress could be one such context; financial hardship has often been found to induce feelings of embarrassment and shame, with one survey finding that 36% of Americans feel embarrassed by their finances.³⁴ Individuals that are ashamed of their finances can be fearful of judgment from others, and therefore may not seek the help that they need.³⁵ Research has indeed shown that shame can lead individuals to disengage from their financial situation—avoiding help, making counterproductive decisions, and ultimately creating a vicious cycle of shame and financial difficulty.³⁶ This raises important questions about whether a financial advice chatbot that clearly presents as non-human, while also containing assurances that consumers’ information is private and secure, could increase the accessibility of financial advice and credit counseling. However, concerns about accuracy as discussed above become especially pertinent in such applications, where advice is often complex and variable from person to person, and the consequences of being wrong are especially high.

Research also raises important questions about the race and gender presentation of advice bots, and how these traits might affect customers’ receptiveness. Some research has found that compared to male-presenting bots, female chatbots are perceived as more human, helpful, and nurturing, though other studies have found no difference between the two.³⁷ There are also questions on how users may respond to chatbots with avatars of different races. Research in other contexts has frequently found a homophily effect—that people are more drawn to and more favorable toward people of the same race as them, whereas one study found that consumers of all races perceived a bot with a Black avatar to be more responsive.³⁸ Regardless, the appearance and style of a bot’s presentation may have implications for which customers feel best-served by the bot, and this is an area in which more research would be warranted.

With the increasing use and complexity of AI chatbots in financial services, these various questions and challenges will become increasingly relevant. AI may create opportunities to reach historically underserved customers, either because they are more comfortable with a chatbot over a human, or because the chatbot can present information in a way that is highly tailored to the specific consumer. Regardless, many consumers will undoubtedly prefer to communicate with a human financial advisor—either in general, or when they have more complex inquiries. These considerations underscore the importance of financial institutions continuing to offer clearly advertised alternatives to chatbots and providing simple and straightforward off-ramps from chat-based channels for users to speak to a human.³⁹

B. Data and technology applications for identity verification, fraud detection, and anti-money laundering

As discussed in Treasury’s cybersecurity report, fraud detection is one of the oldest and broadest examples of machine learning deployment in financial services. Machine learning models are often used to monitor bank account, credit card, or other financial transactions data to flag suspicious patterns, sometimes in combination with more traditional rules-based tools. Some anti-money laundering (AML) compliance programs also make use of ML models, although rules-based tools appear to be more common in AML than in fraud detection.⁴⁰ In light of the nature and volume of the data, rapidly evolving fraud patterns, and other considerations, these ML models are often more complex and are updated more frequently than those used in other financial services contexts, although they are not necessarily fully dynamic in the sense of continuously adjusting themselves based on incoming data without developer initiation and validation prior to deployment.⁴¹ Even very large financial institutions often rely in part on models developed by vendors in this context, given that such platforms can provide insights across multiple institutions at the same time.

ML models are also used for identity verification in initial account opening and application processes to estimate the likelihood that an applicant is who they say they are. These models may analyze, for instance, if recent account openings with the applicant’s social security number or ID are suspicious. Related applications of artificial intelligence include systems that analyze an ID card–typically drivers’ licenses–to detect fake IDs.

While generative AI is not currently in wide use in identity verification or fraud detection processes, several organizations have announced upcoming generative AI-powered products. For example, some FIs and service providers are experimenting with using transformer models (which rely on a model architecture normally used for generative AI applications) for use in transaction monitoring.⁴² There is also interest in using generative AI to create synthetic datasets to train anti-fraud and AML models, although the efficacy of this technique is unclear.⁴³ Public research into the utility and efficacy of techniques and architectures that are used in generative AI for identity verification and transaction monitoring could be helpful, including potential inclusion implications. Some stakeholders argue that financial institutions should instead focus on more basic prevention and detection strategies, such as using two- or multi-factor authentication and refining step-up verification methods, before investing in generative AI for identity verification and transaction monitoring.⁴⁴

Generative AI can also be used by fraudsters to attempt to fool AML and anti-fraud systems, however. As Treasury’s recent report detailed, AI may be used to increase the sophistication of social engineering attacks as well as to generate fake IDs and imitate victims to bypass voice authentication systems.⁴⁵ Artificial intelligence and advanced analytical techniques can be used to detect these imitations with varying degrees of success.⁴⁶ Yet whether these defenses can keep pace and whether safeguards implemented by commercial generative AI providers are having meaningful effects are open questions as AI-enabled fraudsters become more sophisticated.

While stakeholders tend to view identity verification and transaction monitoring almost exclusively through a security lens, these processes have large inclusion impacts that should be monitored and managed carefully as they evolve. An estimated 11.6 percent of unbanked households cite that do not have the personal identification required to open a bank account,⁴⁷ and transaction monitoring models that mistakenly flag transactions as illegitimate can prevent consumers from completing important transactions and, in extreme cases, lead to account closures.⁴⁸ Financially vulnerable consumers can also be disproportionately affected where financial institutions fail to detect fraud, since losses of even small dollar amounts can represent a significant percentage of monthly income or balance sheet savings. The Federal Trade Commission’s fraud surveys have found Black and Latino consumers experience fraud at a higher rate than non-Latino White consumers,⁴⁹ and the Identity Theft Resource Center (ITRC) reports that women and Black victims seek its assistance at higher rates than their proportion of the general population.⁵⁰

Particularly as large data breaches and fraud losses continue to mount,⁵¹ financial system stakeholders are increasingly investing in new types of data, upgrading their monitoring models, and pursuing additional data sharing initiatives. These innovations include broadening identity verification systems to rely on an increasingly diverse range of data, such as crosschecking basic information with wireless carriers and other companies and “digital footprint” data concerning keystroke timing and other device usage patterns, as well as migrating away from rules-based systems toward increasing use of ML predictive models. Stakeholders are also exploring a range of “privacy enhancing technologies” such as encryption and federated machine learning models, to balance the potential benefits of exposing analytical models to more data with the risks to privacy and security from moving and aggregating more customer information.

As identity verification and transaction monitoring become more advanced, the inclusion impacts of these developments–whether positive or negative–should be carefully monitored and considered. The adoption of more diverse data sources and sophisticated analytical techniques could potentially be helpful for consumers who lack extensive traditional financial histories. At the same time, as fraud and illicit finance accelerate, there is a risk that financial institutions may increasingly consider consumers with more limited identity documentation or less traditional financial patterns to be too risky to serve. A forthcoming FinRegLab paper will explore these trends and concerns in greater detail.

C. Back-office applications to increase the nimbleness of model and product development

Use of machine learning and natural language processing in back-office applications often receives less public attention but has potentially important implications for financial inclusion where these tools can help financial institutions be more nimble in processing data, developing new products, or migrating away from legacy systems to platforms that can be used to tailor and deliver financial products and services to historically underserved populations.

One potentially compelling illustration of such use cases concerns the deployment of techniques and model architectures that are commonly used in generative AI applications to help parse information from the text fields on bank account statements that helps to describe the payee or payor and other details of individual transactions. Financial institutions may wish to use this data to help with underwriting and fraud monitoring and building budgeting and other personal financial management tools, among other uses. However, the language of these memo fields is not standardized, and formats vary widely across financial institutions. While categorization models have improved over time, financial institutions are eager for more accurate and reliable techniques.

Financial services providers are using a variety of approaches to convert the data into more structured and standardized formats for analysis. For example, the B2B payments processor Slope has leveraged various models, such as GPT and BERT, to create unsupervised clusters of inflows and outflows, allowing them to see streams of incomes and expenses.⁵² In late 2023 Slope also developed TransFormer, a highly fine-tuned model built on an open source LLM, which can label and categorize transactions with a reportedly high level of accuracy.⁵³ Slope now uses this model in its credit monitoring dashboards, and is planning to integrate it into its underwriting system. Another company working on similar technology is Hyperplane, which was recently acquired by Brazil’s Nubank.⁵⁴ Hyperplane works with banks to convert their various forms of first-party data into a more usable form. Like with Slope, a large component of this includes transaction data. Hyperplane offers services utilizing this data to train and fine-tune predictive models, and to develop marketing campaigns targeting users with financial products.⁵⁵

Such AI technologies could potentially help to derive significant additional insight from bank account data for underwriting, personal financial management, and other applications, as well as to help financial services providers in a variety of other contexts such as mining customer service notes for trend spotting and compliance purposes and migrating their data from one platform to another.⁵⁶ Especially for financial services providers who are reliant on older legacy systems, the ability to migrate to newer platforms could be a significant boost to promoting more innovative uses of data and technology.

Many of these use cases involve converting unstructured data into structured data, often requiring text extraction exercises. As discussed in the background section, companies have a variety of options to choose from depending on the specific task and potential trade-offs between complexity, cost, and accuracy. The accuracy of LLMs can be improved with RAG (Retrieval-Augmented Generation) or fine-tuning processes, although they can be costly and complex to deploy. Simpler traditional models are a potential alternative for specific tasks where the context surrounding a particular word or phrase is less critical.⁵⁷

Next steps to facilitate the adjustment of market practices and regulatory frameworks

While the regulatory frameworks governing financial services have provided a useful starting point in affirmatively managing AI risks that many other sectors do not have,⁵⁸ it is important to consider how both business practices and regulatory expectations may need to evolve to promote consistently responsible, fair, and inclusive implementation of ML/AI applications in financial services. Given the diversity and speed of evolution in technologies, use cases, market practices, and policy debates, this requires particularly careful balancing between the advantages of consistent baseline standards and the need for tailored approaches. Diversity of backgrounds and disciplines is also critical, given that the issues posed by ML/AI are not simply technical in nature but rather implicate a range of broader economic, policy, and dignitary considerations.

As we look ahead to what actions Treasury and other stakeholders can take to promote responsible innovation and competition in the use of AI for financial services and to protect consumers and other potentially impacted parties from unintended consequences, several actions could help the financial services ecosystem move toward more rapid identification and implementation of best practices and regulatory safeguards:

• Increasing resources to support the production of public research, engagement by historically underrepresented and under-resourced actors, and broad intra- and intersector dialogue: As the Treasury Department has highlighted in its own past reports, technology resources and expertise have a significant impact on the extent which stakeholders can adopt ML/AI applications and engage in the process of refining related market practices and policy frameworks. Considering ways to facilitate public research that is specifically grounded in the regulatory context of financial services as well as to support greater engagement by smaller financial services providers, historically underserved communities and their advocates, civil society and academic organizations, and government agencies (both regulators and law enforcement) is critical to ensuring that ML/AI adoption operates to the benefit of broader populations and the general economy rather than narrower groups of providers or customers. For instance, increasing resources for public research that considers the financial inclusion and health implications of particular use cases and to facilitate the creation of high quality datasets for use by the financial services sector for development and testing could help to increase the knowledge base of all stakeholders about evolving technologies and market practices. Additional guidance and supervisory activity by regulators could help to address the particular challenges that smaller providers face in both adopting new technologies and performing due diligence on vendors.⁵⁹ Creating opportunities for a broad range of financial services stakeholders to dialogue about emerging issues and to draw upon broader debates about data science and standards in other sectors is also critically important at a time of rapid change.
• Careful consideration of data governance practices and standards: On a related note, questions regarding data quality, accessibility, and protection are fundamental to both the potential benefits and risks of ML/AI applications. Many of the worst headlines concerning ML/AI applications gone wrong relate to flaws in the nature or treatment of underlying data,⁶⁰ and many of the most promising use cases of ML/AI for financial inclusion also hinge in significant part on the ability to access new data sources.⁶¹ Yet data bias and governance issues also arise in contexts that do not involve machine learning in the first instance, and therefore require continuing direct attention in their own right to further refine best practices and regulatory expectations. While federal law provides more detailed and robust protections for consumer financial data than for many other categories of consumer information, many of the consumer financial laws have not had meaningful updates in several decades. They also vary substantially as to what products and services are covered, how they extend to small business owners (if at all), and the level of regulatory monitoring and enforcement.⁶² Consideration of ways that synthetic data and privacy enhancing technologies can potentially facilitate improvements to model accuracy while avoiding the creation of large lakes of personally identifiable information is particularly critical as fraud pressures mount.
• Review of other risk management and customer protection frameworks that apply to automated decisionmaking: As detailed in FinRegLab’s prior research, model risk management expectations and fair lending requirements have pushed financial services providers to manage affirmatively for performance, fairness, and transparency risks when adopting ML/AI in the context of credit underwriting. However, those regulatory regimes do not apply equally to the entire financial services sector, prompting concerns both about concentration of risk and about level playing fields between competitors. Accordingly, some stakeholders have suggested that imposing basic governance expectations on nonbank financial services providers could be beneficial to the broader ecosystem. Stakeholders are also pointing to ways in which the existing regulatory guidance could be updated and expanded to address topics that are becoming more urgent in the ML/AI era, such as standards for evaluating post hoc explainability tools and for evaluating multiple underwriting models to determine whether they constitute a “less discriminatory alternative” for purposes of disparate impact compliance.⁶³ In evaluating the utility of existing frameworks, one important consideration is whether to focus on machine learning and artificial intelligence specifically, on a broader range of statistical models and automated systems, or on the performance of the general function regardless of the extent it is executed by humans, computers, or some combination of the two.⁶⁴ Consistent high-level principles or standards may be particularly useful at this stage of evolution, given the definitional challenges discussed above, the fact that technologies are evolving rapidly, and the universal importance of qualities such as accuracy and fairness.
• Broader efforts to increase opportunity, fairness, and economic participation: It is also critical to note that while filling information gaps and adopting more predictive models could help substantial numbers of consumers and small business owners access more affordable credit, such actions will not by themselves erase longstanding disparities in income and assets or recent hardships imposed by the pandemic. These factors will continue to shape whether and how customers access financial services, for instance by affecting the number of loan applicants who are assessed as presenting significant risk of default, which will in turn continue to affect whether they are granted credit and at what price. This underscores the importance of using many initiatives and policy levers to address the deep racial disparities in income and assets at the same time that stakeholders in the financial system continue to explore and implement promising data and modeling technique innovations. While there is reason to believe that the financial system can enhance its ability to provide fair and inclusive products and services, relying solely on it to address these cumulative, structural issues would produce too little change too slowly.

FinRegLab appreciates the opportunity to comment on the use, opportunities, and risks of AI in the financial services sector and to emphasize the importance of the issues of financial inclusion, equity, and consumer protection as the sector navigates advancements in AI technology. We would welcome an opportunity to engage further on these topics.

Melissa Koide

CEO & Director

FinRegLab

Prior to establishing FinRegLab, Melissa served as the U.S. Treasury Department’s Deputy Assistant Secretary for Consumer Policy. In that role, Melissa helped to build the first government offered preretirement savings product, the myRA. She also established the $5 million Innovation Fund to support research and strategies to improve consumers’ financial health and their access to safe and affordable financial products and services. Melissa has testified before the Senate Banking and House Financial Services Committees, and she has spoken extensively to policy, industry, and consumer advocacy audiences. She is also a member of the New York State Department of Financial Services’ Financial Innovation Advisory Board.

Kelly Thompson Cochran

Deputy Director and Chief Program Officer

FinRegLab

Prior to joining FinRegLab, Kelly helped to stand up the Consumer Financial Protection Bureau, where she served most recently as the Assistant Director for Regulations. In that capacity, she oversaw rulemaking and guidance activities under the Dodd-Frank Act, Electronic Fund Transfer Act, and various other federal consumer financial laws. Kelly previously was counsel at WilmerHale, where she advised financial institutions on a wide range of legal and regulatory matters including product development, compliance, enforcement, and litigation. Kelly also conducted research on financial services innovation, community reinvestment, and other topics at the University of North Carolina at Chapel Hill.

Download PDF