Artificial Intelligence and the Economy: Implications for Central Banks

After completing this reading, you should be able to:

• Identify and describe the risks arising from the widespread use of AI applications in the financial sector.
• Describe how central banks can harness AI to fulfill their policy objectives.
• Explain the macroeconomic impact of AI, including implications for firms’ productive capacity and investment, labor productivity, household consumption, economic output, inflation, and fiscal sustainability.
• Explain how the use of AI presents new opportunities and challenges for central banks, including the central banks’ role as users and providers of data, and the trade-offs posed by their use of both internally-developed and external AI models.

Risks Arising from the Widespread Use of AI Applications in the Financial Sector

The integration of artificial intelligence (AI), particularly generative AI (GenAI) and large language models (LLMs), into the financial sector presents both immense opportunities and a complex array of risks. These risks span various dimensions, impacting systemic stability, data security, ethical practices, regulatory compliance, and market integrity. A thorough understanding of these challenges is crucial for financial institutions, regulators, and central banks to ensure the responsible and beneficial integration of AI into the global financial system.

Risks from Widespread AI Use in Finance:

1. Systemic Risk and Financial Stability
   • Homogenization of decision-making and herding behavior: If many financial institutions rely on similar AI models trained on similar datasets or using similar algorithms, their decisions can become highly correlated. This can lead to synchronized market behavior and exacerbate market volatility. For example, during a market downturn, if multiple AI-driven trading systems simultaneously identify the same assets as high-risk and trigger sell orders, this coordinated selling pressure can amplify the downward spiral, potentially triggering a market crash, similar to the “flash crash” events observed in recent years. This herding behavior can destabilize markets and increase the risk of systemic crises.
   • Procyclicality: AI’s capacity for rapid and automated decision-making can intensify economic cycles. During economic expansions, AI-driven credit scoring models might become overly optimistic, leading to excessive lending and fueling asset bubbles. Conversely, during economic downturns, these same models might become overly pessimistic, leading to a sharp contraction in credit availability and exacerbating the downturn, mirroring the credit crunch observed during the 2008 financial crisis. This procyclical behavior can amplify economic fluctuations and increase the risk of financial instability.
2. Data Privacy and Security
   • Data breaches: The large datasets used to train AI models are attractive targets for cybercriminals. A successful data breach could expose vast amounts of sensitive personal and financial information, leading to identity theft, fraud, and reputational damage. The consequences could be comparable to large-scale data breaches like the Equifax incident, which exposed the personal information of millions of individuals.
   • Privacy concerns and inference: AI’s ability to infer sensitive personal details from seemingly innocuous data raises significant privacy concerns. For instance, AI algorithms analyzing transaction data could potentially infer sensitive information about an individual’s health, religious beliefs, or political affiliations, even if such information is not explicitly provided. This raises ethical concerns about data misuse and the potential for discrimination.
3. Bias and Fairness
   • Algorithmic bias: If the data used to train an AI model reflects existing societal biases, the model will likely perpetuate these biases. For example, if a loan application model is trained on historical data that reflects past discriminatory lending practices against certain demographic groups, the model may continue to discriminate against these groups, even if the model itself is not explicitly designed to do so.
   • Lack of diversity in data: AI models trained on homogenous or unrepresentative datasets can lead to unfair or inaccurate outcomes for underrepresented groups. For instance, a facial recognition system trained primarily on images of one race may perform poorly when used to identify individuals of another race. In finance, this could lead to biased credit scoring, insurance pricing, or fraud detection.
4. Model Risk
   • Opacity and lack of explainability (black box problem): Many advanced AI models, especially deep learning models, are “black boxes,” meaning their internal decision-making processes are opaque and difficult to understand. This lack of transparency can make it difficult to identify the factors driving a model’s output, hindering accountability and making it challenging to identify and correct errors or biases. This lack of explainability can also create regulatory challenges, as financial institutions are often required to explain their decision-making processes.
   • Overfitting and generalization issues: AI models can overfit to the training data, performing well on that specific dataset but poorly on new, unseen data. This can be particularly problematic in finance, where market conditions and data distributions are constantly changing. For example, an AI trading algorithm trained on historical market data might perform well during periods of stable market conditions but fail dramatically during periods of high volatility or market shocks.
5. Operational Risks
   • Dependence on AI and human oversight: Over-reliance on AI without adequate human oversight can lead to significant errors and financial losses. The Knight Capital incident, where an algorithmic trading error resulted in a $440 million loss in minutes, serves as a stark reminder of the potential consequences of inadequate human oversight.
   • Cybersecurity vulnerabilities: AI systems introduce new cybersecurity vulnerabilities, such as data poisoning attacks, where attackers manipulate the training data to influence the model’s behavior. For example, an attacker could inject malicious data into a fraud detection system’s training data, causing the system to misclassify fraudulent transactions as legitimate.
6. Ethical and Regulatory Challenges
   • Regulatory compliance gaps: The rapid pace of AI development can outpace the development of appropriate regulatory frameworks, creating compliance gaps and uncertainty for financial institutions. For example, existing regulations regarding data privacy or consumer protection may not adequately address the unique challenges posed by AI systems.
   • Ethical concerns and accountability: The use of AI in decision-making raises important ethical questions about accountability and responsibility. If an AI system makes a decision that has negative consequences, who is responsible? The developers of the model? The users of the model? The institution that deployed the model? These questions require careful consideration and the development of clear ethical guidelines and accountability frameworks.
7. Market Integrity
   • Manipulation risks: AI can be used to manipulate markets by generating fake news, spreading misinformation, or engaging in sophisticated trading strategies designed to exploit market inefficiencies. This can undermine market integrity and erode public trust in the financial system.
   • Information asymmetry: If only a few institutions have access to and expertise in using advanced AI tools, this can create information asymmetry and give these institutions an unfair advantage in the market. This can lead to increased market concentration and potentially create opportunities for market manipulation or insider trading.
8. Market Concentration
   • Third-Party dependency and concentration of AI capabilities: The development and deployment of advanced AI models require significant investment in data, computing power, and specialized expertise. This can lead to a concentration of AI capabilities in a few large technology companies or financial institutions, creating dependencies and potential systemic risks. If these key providers experience disruptions or if their models contain flaws, the impact could ripple through the entire financial system.

Key Takeaways

• AI applications in finance pose risks to both financial stability and cybersecurity.
• Reliance on similar AI algorithms increases systemic risks like market volatility and procyclicality.
• AI can enhance but also pose challenges to cybersecurity, necessitating careful implementation and oversight.
• Mitigating these risks requires a balanced approach, leveraging AI’s strengths while guarding against its potential weaknesses.

Tip: When assessing AI risks, consider both the technological limitations and the socio-economic impacts on financial systems.

How Central Banks Can Harness AI to Fulfill Their Policy Objectives

Central banks play a crucial role in maintaining economic and financial stability, primarily through monetary policy and oversight of the financial system. AI offers powerful tools to enhance these core functions, impacting information collection, analysis, and policy implementation.

Risks from Widespread AI Use in Finance:

1. Enhanced Information Collection, Statistical Compilation, and Analysis
   • Processing vast and diverse datasets: AI, especially machine learning, can process significantly larger and more diverse datasets than traditional statistical methods. This includes structured data like macroeconomic indicators (GDP, inflation, unemployment) and financial market data (stock prices, interest rates, exchange rates), as well as unstructured data like news articles, social media sentiment, satellite imagery, and even textual data from company earnings calls. This capability allows for a more holistic and nuanced understanding of the economic landscape. For instance, AI could analyze
   • Real-time insights (nowcasting): AI can analyze high-frequency data (e.g., credit card transactions, electricity consumption, traffic patterns) to provide real-time estimates of current economic activity (nowcasting). This allows for more timely assessments of economic conditions compared to traditional lagging indicators. For example, by analyzing real-time credit card transaction data, a central bank could get an early indication of changes in consumer spending patterns, which could inform monetary policy decisions. The BIS report specifically mentions using LLMs fine-tuned with financial news to extract sentiment indices from various sources, which can then be used to nowcast financial conditions. Imagine an LLM analyzing news articles and social media posts about supply chain disruptions to provide a real-time assessment of potential inflationary pressures.
   • Granular data analysis: AI enables analysis at a much more granular level, tracking developments across different industries, regions, and even individual firms. This detailed view can reveal specific sectors or regions driving broader economic trends, allowing for more targeted policy interventions. For example, AI could analyze job postings data to identify skill shortages in specific industries or regions, which could inform government policies on education and training. The BIS report suggests using data from job postings or online retailers to track wage and employment dynamics across occupations, tasks, and industries, and satellite data or electricity consumption to monitor regional economic activity.
   • Improved inflation analysis: AI can analyze a greater number of input variables and identify complex, non-linear relationships between factors contributing to inflation, providing a more nuanced understanding of inflationary pressures, especially during periods of rapid change. For instance, AI could analyze the impact of supply chain disruptions, energy prices, and wage growth on inflation in a more holistic way than traditional econometric models. The BIS report describes the use of neural networks to decompose aggregate inflation into various sub-components, accounting for possible non-linearities.
2. Improved Financial Stability Monitoring Supervision
   • Enhanced risk identification and monitoring: AI can analyze vast amounts of data, including supervisory data, transaction data, and market data, to identify emerging risks and vulnerabilities within the financial system more efficiently. For example, AI algorithms can be used to detect anomalies in trading patterns that may indicate market manipulation or to identify patterns of interconnectedness between financial institutions that could amplify systemic risk. The BIS report cites Project Aurora, which used AI to detect simulated money laundering activities by analyzing payment relationships.
   • Early warning systems and stress testing: AI can be used to develop more sophisticated early warning systems for financial crises by identifying subtle patterns and anomalies in financial data that may precede a crisis. Moreover, AI can generate more realistic and comprehensive stress test scenarios by creating synthetic data that reflects a wider range of potential shocks and market conditions. For example, AI can be used to generate synthetic data representing extreme market events, such as a sudden increase in interest rates or a sharp decline in asset prices, to assess the resilience of financial institutions to these shocks.
   • Automated regulatory reporting and compliance: AI can automate regulatory reporting processes and enhance compliance monitoring, reducing the burden on financial institutions and improving the efficiency of regulatory oversight.
3. Enhanced Payment System Oversight and Development
   • Improved efficiency and security: AI can improve the efficiency and security of payment systems by automating processes, detecting fraud, and enhancing KYC/AML compliance. Project Agorá, highlighted in the BIS report, demonstrates how tokenization and AI can streamline cross-border payments by combining messaging, account updates, and clearing into a single atomic operation. This reduces processing time and costs while enhancing security.
   • Fraud detection and prevention: AI can be used to detect and prevent fraud in payment systems by identifying suspicious transactions and patterns of fraudulent activity. For example, AI algorithms can analyze transaction data in real time to identify anomalies that may indicate fraudulent activity, such as unusual transaction amounts, locations, or frequencies.
4. Enhanced Cybersecurity
   • Threat detection and response: AI can be used to enhance cybersecurity by automating routine tasks, improving threat detection, and accelerating response times to cyberattacks. For example, AI can be used to analyze network traffic to identify suspicious activity that may indicate a cyberattack. The BIS report mentions a survey of central bank cyber experts who believe GenAI offers more benefits than risks in enhancing cybersecurity management. Project Raven is cited as an example of AI being used to improve cyber resilience.

Key Takeaways

• Data is a critical resource for AI applications in central banking, and the payment system is a particularly valuable source of data.
• AI can be used to improve nowcasting by leveraging unstructured data and overcoming the limitations of traditional methods.
• AI can enhance the understanding of factors contributing to inflation by analyzing more granular data and non-linear relationships.
• AI can support financial stability analysis by identifying patterns in large datasets and facilitating cross-border collaboration.
• AI can enhance payment system oversight and development by improving efficiency, security, and fraud detection.

• Central banks need to address challenges related to data availability, data structuring, and the integration of AI tools into their operations.

Tip: Remember that AI in central banking is best applied when data is plentiful, and infrastructure supports cross-border cooperation.

The Macroeconomic Impact of AI

AI is poised to have a profound impact on the macroeconomy, affecting various key variables.

1. Firm Productive Capacity and Investment
   • Increased productivity: AI has demonstrated the potential to significantly increase productivity, particularly in tasks requiring cognitive skills. Studies have shown substantial productivity gains for customer support agents, writers, and software developers using AI tools. The BIS report, citing a study with Ant Group, notes that productivity gains are often immediate and largest among less experienced staff. These gains translate to an expansion of firms’ productive capacity, allowing them to produce more output with the same or fewer inputs.
   • Spurring innovation: Beyond direct productivity enhancements, AI can also spur innovation by improving the efficiency of cognitive work, which is crucial for generating new ideas and technologies. This indirect effect on innovation can lead to further productivity growth in the future.
   • Increased investment: Firms are already investing heavily in AI technology, including IT infrastructure and the integration of AI models into their operations. This investment contributes to aggregate demand and can further boost economic growth. The BIS report cites substantial global spending on AI and surveys indicating that many firms prioritize AI in their budgets. Improved prediction capabilities due to AI adoption can further encourage investment by reducing uncertainty and enabling better decision-making. However, the report also acknowledges that AI could introduce new uncertainties that might partially offset this positive impact.
2. Labor Productivity
   • Direct productivity gains: As mentioned above, AI can directly increase labor productivity by augmenting human capabilities in various tasks. The BIS report provides examples of significant productivity increases in specific tasks and occupations.
   • Creation of new tasks and displacement of workers: AI can both create new tasks and make existing ones obsolete. The net effect on labor demand and wages will depend on the relative strength of these two forces. If AI acts as a general-purpose technology, increasing total factor productivity across industries, it could lead to increased labor demand and wage growth. However, if AI primarily automates existing tasks, it could lead to worker displacement and downward pressure on wages, potentially increasing economic inequality. The BIS report highlights the potential for differential impacts across job categories, with some workers benefiting significantly while others face displacement.
3. Household Consumption
   • Reduced search frictions and improved matching: AI can reduce search frictions for consumers by improving product and service discovery and matching. This can lead to increased consumption.
   • Impact of labor market effects: AI’s impact on household consumption will also depend on its effects on labor markets. If AI leads to widespread job displacement and wage stagnation, it could negatively impact consumer spending. Conversely, if AI leads to increased labor demand and wage growth, it could boost consumer spending. The BIS report discusses the potential for AI to affect economic inequality, which would also have implications for consumption patterns.
4. Economic Output
   • Increased aggregate supply and demand: AI’s impact on productivity, investment, and consumption will affect both aggregate supply and demand, leading to changes in economic output. The BIS report illustrates how a permanent increase in productivity due to AI will raise aggregate supply, while increased consumption and investment will raise aggregate demand. The net effect on output will depend on the relative strength of these effects and how expectations are formed.
5. Inflation
   • Short-term and long-term effects: AI’s impact on inflation is complex and depends on expectations and the balance between supply and demand effects. Initially, AI can have a disinflationary impact if productivity gains outpace demand growth. However, if households and firms anticipate future gains from AI and increase consumption and investment accordingly, the initial impact could be inflationary. Over time, as economic capacity expands and wages rise, demand for capital and labor will increase, potentially leading to higher inflation. The BIS report suggests that the initial impact is more likely to be disinflationary, based on the historical experience of previous general-purpose technologies.
   • Impact on price formation: Increased use of algorithmic pricing by firms, facilitated by AI, can lead to greater uniformity and flexibility in pricing, potentially increasing the pass-through of aggregate shocks to local prices and altering inflation dynamics. The BIS report points out that this effect could be influenced by the degree of competition in the AI model and data market.
6. Fiscal Sustainability
   • Impact on debt burden: AI-induced productivity and growth can lead to a reduced debt burden. However, the potential need for fiscal programs to manage AI-induced labor relocation or higher unemployment, combined with potentially higher interest rates associated with faster growth, could moderate this effect. The BIS report concludes that the AI growth dividend is unlikely to fully offset the spending needs that may arise from other factors, such as the green transition or population aging.

How the Use of AI Presents New Opportunities and Challenges for Central Banks

The advent of AI presents both significant opportunities and distinct challenges for central banks, profoundly impacting their roles as data users, compilers, and disseminators, and necessitating careful consideration of trade-offs in model development. AI offers the potential to significantly enhance central banks’ ability to fulfill their mandates of price and financial stability. As discussed previously, AI can improve economic forecasting, financial stability monitoring, payment system oversight, and cybersecurity. Data is a crucial resource for AI, and central banks, with their access to vast amounts of economic and financial data (especially from payment systems), are well-positioned to benefit. AI can unlock the value of this data by identifying patterns, trends, and anomalies that would be difficult or impossible to detect with traditional methods. The BIS report emphasizes the increasing value of data in the age of AI and the importance of appropriate data governance frameworks. AI can also automate routine tasks, such as data collection, processing, and reporting, freeing up central bank staff to focus on more complex analytical and policy-related work. Moreover, AI can facilitate collaboration and knowledge sharing among central banks. This includes sharing data, best practices, AI tools, and even trained models, reducing costs and improving efficiency. The BIS report strongly advocates for a “community of practice” among central banks to facilitate this collaboration.

However, central banks also face several challenges in adopting AI:

• Data availability, quality, and governance are essential for effective AI implementation. Central banks face challenges in accessing and integrating diverse data sources, particularly unstructured data held by the private sector. Robust data governance frameworks are crucial to ensure data quality, privacy, and security. The BIS report highlights the rising costs and tighter use conditions imposed by commercial data vendors.
• A key trade-off arises between using “off-the-shelf” (external) AI models and developing in-house (internal) models.
  • External models offer cost-effectiveness and leverage the expertise of private sector companies, especially in the short run. However, they come with reduced transparency (the “black box” problem), increased dependence on external providers, and potential risks related to data security and vendor lock-in. The BIS report explicitly mentions the risks associated with market concentration among AI providers.
  • Internal models offer greater control, transparency, and customization, allowing central banks to tailor models to their specific needs. However, they require significant investment in IT infrastructure, computing power, and specialized expertise (data scientists, AI engineers).
• Implementing and managing AI systems requires specialized skills in areas such as data science, AI engineering, and cybersecurity. Central banks face challenges in attracting and retaining talent in these competitive fields, especially given the higher salaries offered by the private sector. The BIS report highlights this as a top concern for central banks.
• It is also important to remember that AI is not a substitute for human judgment. It requires supervision by experts with a deep understanding of macroeconomic and financial processes. The BIS report emphasizes that AI models used for financial stability monitoring still require human oversight.
• Finally, AI systems introduce new cybersecurity vulnerabilities, such as data poisoning and prompt injection attacks, which can compromise data integrity and model performance.

Central banks play multifaceted roles concerning data. They are extensive users of data for policy analysis, decision-making, and communication. AI enhances their ability to analyze this data and extract valuable insights. Central banks also have a long history of being compilers of data, either collecting it themselves or drawing on other official and commercial sources. This role becomes even more important in the age of AI, as high-quality data is essential for training effective models. Finally, central banks also act as providers of data to other government agencies and the public, contributing to transparency and informed decision-making. This role can be enhanced through the use of standardized metadata and data sharing initiatives, as advocated in the BIS report.

Central banks can mitigate the trade-offs associated with AI adoption through collaboration and knowledge sharing. This includes pooling resources and knowledge, sharing data and best practices, and developing common standards. The BIS report highlights the benefits of cooperation in areas such as data acquisition, staff training, and model development. Robust data governance frameworks are essential to ensure data quality, privacy, security, and interoperability. This includes establishing clear guidelines for data collection, storage, use, and sharing. The BIS report emphasizes the importance of metadata and the adoption of standards such as SDMX to facilitate data sharing and interoperability. By addressing these challenges and leveraging the opportunities through collaboration and sound data governance, central banks can effectively harness the power of AI to enhance their core functions and contribute to economic and financial stability.

Question

A key trade-off for central banks considering AI adoption involves choosing between “off-the-shelf” (external) and in-house (internal) models. What is a primary disadvantage of using external AI models?

A) Higher development costs and longer implementation times.

B) Reduced transparency and potential vendor lock-in.

C) Limited access to large datasets for model training.

D) Difficulty in attracting and retaining specialized AI talent.

Correct Answer: B

External “off-the-shelf” AI models are often proprietary, meaning central banks may have limited visibility into how these models function (lack of transparency) and may be reliant on the vendor for updates, maintenance, or modifications (vendor lock-in). This can hinder the institution’s ability to fully understand and trust the model’s outputs, which is particularly concerning in critical areas like monetary policy and financial regulation.

A is incorrect: These are disadvantages of internal model development.

C is incorrect: External models often benefit from the large datasets held by the providers.

D is incorrect: This challenge pertains to in-house models, as off-the-shelf solutions do not require the institution to have extensive AI expertise.