RatioLogo
Back

Measuring a Supersonic Jet with a Wooden Ruler

What if the gatekeepers of the global financial system are trying to measure a supersonic jet with a wooden ruler?

For decades, banks have managed risk through committee meetings and manual spreadsheets—a "checkpoint" system designed for simple math. But as Artificial Intelligence begins to run everything from fraud detection to asset pricing, this legacy infrastructure is hitting a breaking point.

The Crumbling Legacy System

A Stark Mismatch in Speeds

A new technical proposal reveals a critical misalignment. While AI model parameters are increasing 10x annually, the human-led governance processes meant to oversee them can take up to 12 months to complete. By the time a risk officer signs off, the model is often already obsolete.

The Scale of the Problem

A Systemic Threat

This isn't just a headache for bankers; it is a systemic threat to the economy. The COVID-19 pandemic served as a "natural stress test" that saw most financial AI models fail because they couldn't recalibrate fast enough under manual constraints.

The Staggering Costs of Non-Compliance

  • Between 2008 and 2016, regulatory failures cost the industry USD 320 billion in fines.
  • The cost of staying compliant is skyrocketing, hitting USD 80 billion in 2019.
  • This spending is projected to reach USD 120 billion by 2025.

The Proposed Solution: Self-Regulating AI

The researchers argue that the era of "parameter-level" human review is over. You cannot manually audit a model with trillions of moving parts. Instead, they propose a shift to "Self-Regulating AI."

The "Governance Block" Framework

This new framework suggests embedding "governance blocks" directly into the AI software. Rather than waiting for a committee review once a year, these blocks would monitor fairness, bias, and accuracy in real-time, acting like an automated flight correction system for high-speed algorithms.

The Urgency and the Obstacles

Who Needs It Most?

The urgency is highest among the most sophisticated players. Data shows that 94% to 97% of AI-mature financial firms view regulatory complexity as their primary obstacle. These institutions are trapped in a loop where the "model aging" caused by slow reviews creates more risk than the AI itself.

The Roadblocks to Implementation

However, the path to automation is not without significant hurdles:

  1. Lack of Validation: The authors acknowledge their proposed system-level framework lacks empirical validation from a live, large-scale implementation.
  2. The Human Bottleneck: Even a self-regulating system requires a "human-in-the-loop" for final, high-stakes decisions, which could remain a bottleneck.

Conclusion

For now, the industry remains in a precarious transition, waiting for global regulators to catch up to a reality where the machines are evolving faster than the rules can be written.

Source: Towards Self-Regulating AI: Challenges and Opportunities of AI Model Governance in Financial Services. Eren Kurshan, Hongda Shen, and Jiahao Chen. Presented at the ACM International Conference on AI in Finance (ICAIF ’20).