Scrutiny of AI Bias in Finance Grows
As AI models are embedded into core banking functions, concerns over their potential to perpetuate social group biases are increasing. Recent research highlights that large language models in finance can amplify harmful biases if not carefully governed. This is raising the bar for explainability and auditability in AI-driven fraud and risk management systems, with banks and regulators expected to demand greater transparency.
- Research from Lehigh University using OpenAI's GPT-4 Turbo found that Black mortgage applicants needed credit scores about 120 points higher than white applicants to achieve the same loan approval rate. This highlights how AI can perpetuate historical biases, a risk noted by the Consumer Financial Protection Bureau, which has expanded its definition of "unfair" practices to include discriminatory AI conduct. - AI bias often originates from the data used to train the models, which can reflect past discriminatory practices like redlining. For example, a major credit card issuer's AI system allocated lower credit limits to women than to men with identical financial profiles because the historical data reflected this bias. - Proxies for protected characteristics, such as ZIP codes or educational history, can inadvertently introduce bias into lending algorithms. A study by the National Bureau of Economic Research showed that even after accounting for creditworthiness, mortgage algorithms charged higher interest rates to Black and Hispanic borrowers. - Regulatory bodies are increasing their focus on AI fairness, with the EU's AI Act classifying credit assessment as "high-risk" and requiring compliance within 24 months. In the U.S., financial institutions must adhere to the Equal Credit Opportunity Act (ECOA), which forbids lending discrimination. - To combat bias, some financial institutions are adopting regular algorithmic audits and utilizing diverse testing datasets. Singapore's Veritas Toolkit and AI Verify are examples of operational frameworks being used to establish measurable, use-case-specific fairness standards. - The "black box" nature of some complex AI models makes it difficult to understand their decision-making processes, posing challenges for transparency and regulatory compliance. This lack of explainability complicates efforts to prove that a model's decisions are fair and unbiased. - Digital identity verification is a key tool in fraud prevention, utilizing technologies like biometric authentication and AI-driven behavioral analysis to secure payments. These systems help establish trust and reduce the risk of fraudulent activities in real-time payment environments. - The U.S. real-time payments landscape is rapidly growing, with the FedNow service reaching over 1,600 financial institutions since its 2023 launch and processing $853.4 billion in 2025. It joins the RTP network, which handled over $1.3 trillion in 2025, signaling a broader shift toward instant payment infrastructure.
