AI is Entering Regulated Territory: Understanding the Supervisors' Perspective for Model Justifiability in Financial Crime Detection

Title of the Paper

AI is Entering Regulated Territory: Understanding the Supervisors’ Perspective on Model Justifiability in Financial Crime Detection

Paper Information

• Subject Area: Application of Artificial Intelligence (AI) in financial crime detection, particularly in the domain of Anti-Money Laundering and Counter-Terrorism Financing (AML-CFT), focusing on model transparency and compliance.
• Keywords: Model interpretability, anti-money laundering, highly regulated environments, financial regulation, artificial intelligence, model justifiability, transparency, algorithmic fairness, AI regulation, regulatory compliance

Research Background and Problem Statement

• Identified Challenges: Although Artificial Intelligence (AI) technology has significant potential in highly regulated industries such as banking and finance, its actual adoption rate remains relatively low. Specifically, in the AML-CFT domain, the use of AI technologies raises concerns about transparency and trustworthiness due to the "black-box" nature of the models.
• Significance of the Research Problem: Current rule-based AML-CFT systems are inefficient, with many criminal activities going undetected, and less than 1% of illicit funds being traced or frozen. At the same time, AI has immense potential to reduce labor costs, uncover unidentified patterns, and improve detection accuracy. Therefore, addressing the transparency issues of AI systems is crucial to promoting their large-scale adoption.
• Research Motivation and Related Work: While there has been research on AI applications in AML-CFT, there is a lack of methodological studies focusing on model transparency and compliance, particularly from the perspective of regulators. A deeper understanding of how regulators perceive AI model interpretability and its application within complex legal frameworks is needed.

Solution

• Methods and Steps:

  1. Investigate how regulators supervise and audit anti-money laundering activities within socio-technical-legal contexts.
  2. Explore regulators' requirements for AI system justifiability through scenario-based workshops.
  3. Conduct qualitative legal research to clarify the specific transparency requirements imposed by AML-CFT regulations on AI.
  4. Validate the findings from legal analysis and demand summaries through expert interviews and field data collection.

• Innovations:

  1. Identified seven core requirements from regulators for AI systems, including model efficiency measurement, local sample analysis, and validation of appropriate use.
  2. Introduced a scenario-based design approach that integrates human-computer interaction (HCI) and legal analysis to provide more specific, user-oriented recommendations for AI in regulatory contexts.

• Technologies Used:

  1. Qualitative research methods: Scenario-based workshops to gather feedback on regulators' transparency requirements for AI.
  2. Legal and regulatory analysis: Parsing AML-CFT regulatory content and its compatibility with AI applications.

Research Findings

• Specific Findings:

  1. Described the socio-technical-legal supervisory system and common auditing methods used by regulators in AML-CFT.
  2. Clarified the transparency requirements imposed by AML-CFT regulations on AI systems, such as adapting to risk classification, monitoring anomaly detection capabilities, reporting quality, and human-machine collaboration.
  3. Extracted seven key requirements from regulators for AI systems (e.g., the need for understanding model foundations, global efficiency measurement, and judgment of faulty samples).
  4. Highlighted the potential and limitations of explainable artificial intelligence (XAI) in supporting legal compliance.

• Advantages Compared to Existing Solutions:

  • Provides a deeper analysis of AI transparency requirements from the regulatory perspective, addressing gaps in existing research on legal-oriented needs assessment.
  • Combines legal analysis and HCI design methodologies to propose a multidisciplinary framework for studying AI compliance.

• Experimental and Evaluation Results:

  • Identified challenges posed by the "opacity" of AI to AML-CFT compliance work among participating regulators, while recognizing that sample analysis and global evaluations of model operations can mitigate these challenges.
  • Offered specific design suggestions for conventional explanation methods such as feature importance analysis and graph network visualization, while highlighting potential risks of misuse (e.g., confirmation bias).

• Limitations and Future Directions:

  1. Limitations:
     • The study design is limited to AML-CFT regulations and specific scenarios (primarily focusing on model scoring and anomaly detection); other use cases are not covered.
     • Participant feedback is based on preliminary understanding, and requirements may evolve as AML-CFT technology advances.
  2. Future Directions:
     • Develop comparative explanation tools tailored to regulatory scenarios to support justifiability analysis of erroneous cases.
     • Enhance the reliability of global evaluation metrics for AI systems, particularly by developing new anti-discrimination and transparency testing methods in the AML domain.
     • Explore how integrating Financial Intelligence Units (FIUs) can provide deeper insights into the practical effectiveness of AML-CFT.

Conclusion

This study combines legal and human-computer interaction methods to establish a clear framework of requirements for AI compliance in AML-CFT. Future research in related fields can leverage these findings to optimize the transparency and interpretability of AI decision-making systems, thereby enhancing their acceptance and legitimacy in highly regulated environments.