Artificial Intelligence and Bank Supervision

After completing this reading, you should be able to:

• Describe historical evolution and common types of AI-based applications used in the financial sector.
• Explain the advantages of implementing AI-based applications to the banking services companies and their customers.
• Discuss the disadvantages and difficulties for financial companies using AI.
• Clarify the specific issues faced by banks and regulators arising from utilizing AI in modeling and valuation.

Historical Evolution and Types of AI in the Financial Sector

The Inception of AI and Its Growth in Finance

Artificial Intelligence (AI) as a concept dates back to the mid-20th century, originating from the thoughts of influential figures such as mathematician and cryptographer Alan Turing. Turing sparked the AI discussion with his 1950 essay on the potential of machines to imitate human intelligence. This intellectual curiosity gave rise to significant AI milestones, such as IBM’s Deep Blue defeating chess champion Garry Kasparov in 1997. Since then, AI has experienced a surge in growth and capabilities, primarily through the advancement of machine learning (ML) methodologies, including the development of digital neural networks. These neural networks have become prevalent in classifying various types of data, including text and images, which has facilitated AI implementation in numerous industrial applications and consumer services—ranging from Google searches to Netflix recommendation algorithms. In recent developments, ML has become fundamental to the creation of innovative generative AI applications, like ChatGPT. These programs are engineered to perform various tasks, including engaging in productive dialogues with people. This advancement is particularly notable in the context of U.S. financial institutions.

AI’s Foray into the Financial Sector

The U.S. financial industry has actively engaged with AI technology, applying it for a multitude of functions. Despite this, the sector has exhibited a cautious and measured approach to full AI integration. In a 2019 study, McKinsey & Co. discovered that the adoption of AI in the financial services industry was quite limited. Only about 36% of those surveyed said their companies used AI to automate back-office tasks. The usage of AI in customer service chatbots and for detecting fraud or assessing creditworthiness was even less common, at 32% and 25%, respectively. Cornerstone Advisors, another consultancy, conducted a similar survey in 2022 among bank and credit union leaders. Their findings were even more striking, revealing that a mere 25% had implemented AI for process automation, and just 18% had introduced AI-based chatbots. The moderate uptake suggests an industry that is still in the process of comprehensively embracing AI’s potential.

The Spectrum of AI Applications in Finance

Financial institutions in the U.S. have developed and deployed AI for various purposes, signaling a steady growth and potential expansion in the role of AI within the sector. Awareness of this trajectory has prompted bank regulators to keep a keen eye on how banks use AI. This oversight has materialized in actions such as the Request for Information (RFI) issued in March 2021 by key bank regulators like the Comptroller of the Currency (OCC) and the Fed Board of Governors, which sought to gather more insights into the banks’ current and future practices related to AI. It elucidates a commitment to adapt regulatory measures as AI continues to evolve within the landscape of the financial industry.

Advantages of AI Implementation in Banking

• Enhanced efficiency and customization: In the financial sector, adopting AI technology has led to substantial improvements in operational efficiency and service customization. Banks deploy machine learning (ML) to streamline back-office tasks, enhance customer service through chatbots, and increase the precision of fraud detection and credit evaluations. The introduction of AI allows for sophisticated data analysis, leading to the identification of fraudulent transactions and minimizing false positives.
• Proactive Problem solving and improved customer interaction: AI-based applications within banking operations facilitate proactive problem-solving by highlighting suspicious patterns that warrant further investigation. Tools employing AI provide banks with the capacity to monitor transactions dynamically, aiding in the prevention of financial crimes with their enhanced detection capabilities. Moreover, customer interactions are significantly improved through AI with services such as efficient chatbots that can handle a high volume of queries, leading to a better overall customer experience. In addition, AI systems utilizing pattern recognition have enabled clients to remotely deposit checks, eliminating the need for additional visits to physical bank branches.
• Credit evaluation: AI applications are revolutionizing credit evaluation and loan underwriting, particularly within the retail sector, including credit card and automobile financing. These AI systems provide enhanced efficiency and accuracy in credit decisions while ensuring compliance with legal standards like the Equal Credit Opportunity Act. Despite their complexity and perceived opaqueness, efforts are being made to explain AI-driven credit decisions, maintaining a balance between innovative AI models and traditional credit evaluation methods used by banks and credit bureaus.

Disadvantages and Challenges in AI Use for Financial Companies

• Explainability and transparency issues: A critical concern surrounding the use of AI within banking is the explainability and transparency of AI-derived decisions and models. Large AI models based on machine learning algorithms can be opaque, lacking the clear-cut variables and interpretable coefficients present in conventional statistical models. This lack of transparency makes it difficult to discern the reasoning behind AI system outputs, which is particularly problematic for AI-based credit models that may unwittingly perpetuate biases such as racism, contravening federal anti-discrimination laws. A lack of model explainability can hinder banks’ ability to comply with various regulations, including consumer protection requirements.
• Regulatory compliance and bias mitigation: The difficulty in interpreting AI models also poses significant challenges in adhering to regulatory requirements, including consumer protection mandates. Bank regulators emphasize the necessity of comprehending and explaining how AI models reach conclusions to ensure that these systems comply with applicable laws and regulations. An overwhelming concern is that AI systems might reflect or amplify biases in their training data, thereby inadvertently entrenching discriminatory practices.
• Evolving fraudster methods: AI models are commonly used in banks for fraud detection, particularly in credit card fraud and identity theft cases. However, fraudsters continuously innovate their attack methods, leading to an escalating game of cat and mouse between them and the banks.
• Complexity of AI models: Traditional credit evaluation models, similar to those used by national credit bureaus, are still employed by most banks. These models typically use a single formula based on a small group of indicators. In contrast, AI models involve multiple layers of complex analysis with numerous quantitative and qualitative inputs, making them more difficult to understand and interpret.

Issues in AI for Banking Modeling and Valuation

Model Validation, Maintenance, and Data Reliability

As AI models adapt and learn from new data over time, they necessitate continuous monitoring for model validation, maintenance, and documentation. This includes vigilance to prevent the incorporation of indicators that function as proxies for prohibited attributes like race in credit evaluations and other decision-making processes, potentially violating anti-discrimination laws. Bank regulators emphasize the need for banks to understand the inner workings of their AI models and explain their results to ensure compliance with laws and regulations. The concern is not only the difficulty in maintaining AI models but also ensuring the reliability and quality of the data feeding into these models, given that AI’s performance is heavily dependent on the data it processes. Banks are exploring “nontraditional” data sources for AI models, such as rent, utility payments, and cash flow patterns. This raises concerns about compliance with consumer protection laws. Questions remain about the effectiveness and privacy implications of such data.

Regulatory Engagement and Prudent Model Evaluation

Regulators express a carefully balanced optimism regarding the potential for AI to enhance certain credit evaluation aspects. The approach taken by banks involves rigorously testing new models alongside existing ones before making informative decisions based on the new systems. The broader view is managing the integration of AI technologies thoughtfully, with an awareness of their possibilities and pitfalls. Bank supervisors are forward-looking, acknowledging both the apprehensions and the prospective upsides of leveraging AI in banking operations.

Practice Questions

Question 1

In the early stages of integrating AI into its operations, the U.S. financial sector anticipates potential benefits and challenges. The richness of AI applications has contributed to increased operational efficiency and customer satisfaction. However, one of the main concerns arising from the use of AI for credit modeling is the technology’s opacity and potential to unintentionally perpetuate human biases. Which of the following best describes the concern expressed by bank regulators regarding AI-based credit models?

1. AI models always produce less accurate credit scoring than traditional methods do.
2. AI models’ inherent complexity and inability to explain their outputs may unintentionally violate anti-discrimination laws.
3. The cost of implementing AI is prohibitively high, outweighing its benefits in credit evaluations.
4. AI technology lacks the capability to learn and adapt over time, making it unsuitable for dynamic credit modeling.

Correct Answer: B)

AI’s complexities can make it hard to explain how the system arrives at its results. The concern is that such opaqueness could lead to unintentional algorithmic biases, which might conflict with federal anti-discrimination laws aimed at preventing injustices such as racism.

A is incorrect. It does not address the issue of potential legal violations due to biases or lack of transparency.

C is incorrect. The primary concern noted is not the cost but the explainability and potential biases of AI models.

D is incorrect. AI technology is actually known for its ability to learn and adapt, which is not the primary concern raised by regulators.

Question 2

Bank supervisors have conveyed optimism about the technology’s potential benefits despite recognizing numerous potential pitfalls of banks’ use of AI-based applications. Concerns have been raised regarding the ongoing maintenance of AI-based credit models as these continuously evolve by learning from new data, which could present challenges for ensuring legal and ethical compliance. In response to an interagency Request for Information (RFI), financial institutions have discussed their data monitoring processes. Which of the following responses to the RFI best illustrates the industry’s stance on data quality for AI models?

1. Financial institutions highlighted that the data used for AI models is of better quality than that used for traditional models, eliminating concern over biases.
2. They emphasized that the risks associated with poor data quality are unique to AI models and do not affect traditional models.
3. Mortgage industry representatives noted that alternative data sources, which could improve AI model fairness, are not allowed by major federal housing agencies.
4. They stressed that the risks of poor data quality are not unique to AI-based models and that banks use consistent data monitoring processes across all models.

Correct Answer: D)

The response by financial institutions to the RFI suggests that the issues associated with data quality affect both AI and traditional models alike. They argue that the processes for monitoring data used in AI models are consistent with those used in traditional ones, implying a uniform approach to data quality irrespective of the technology deployed.

A is incorrect. There is no claim that data for AI models are of superior quality compared to traditional models.

B is incorrect. The response did not suggest that risks associated with poor data quality are exclusive to AI models.

C is incorrect. Although it states a factual situation within the mortgage industry, it does not represent the broader industry’s stance on the issue of data quality in AI models.