Towards Responsible AI for Financial Transactions

Explainability is critical for financial AI: The paper identifies explainability as a foundational principle for building trust in AI systems used for financial transactions, alongside fairness, privacy, accountability, transparency, and soundness.

Multiple explanation techniques are needed: The researchers used two complementary approaches—Shapley additive explanations (SHAP) for feature importance analysis and a hybrid method combining text clustering with decision trees—suggesting that no single explanation method fully captures how complex financial AI models make decisions.

Understanding model decisions reveals vulnerabilities: By extracting explanations of how the model worked, researchers were able to design targeted attacks that could fool the system, demonstrating that transparency about AI decision-making can expose security weaknesses that need to be addressed.

Financial AI handles diverse data types: The model processes numerical, categorical, and textual information simultaneously for transaction classification, reflecting the complexity of real-world financial data that AI systems must navigate responsibly.

Robustness testing is essential: The paper emphasizes that responsible AI in finance requires not just explaining decisions, but actively testing whether models can withstand adversarial attacks—a critical safeguard before deploying AI in high-stakes financial applications.