The Bank of Russia proposed regulating AI models used by credit organizations. The Central Bank considers it dangerous when dozens of banks operate on one global platform but cannot verify what data it was trained on and why it makes certain decisions.
The problem is that an error in one system can quickly become widespread. Banks can simultaneously lower limits, incorrectly assess borrowers, or direct money into identical assets, amplifying market movements. According to Timur Aitov, an expert at the Russian Chamber of Commerce and Industry, in some segments, credit portfolio losses could reach 3–5% per quarter.
There is also a dependence on foreign suppliers. A malfunction, license change, or service shutdown risks disrupting the operations of several market participants at once. Additionally, malicious actors could intentionally distort input data, and banks would have to wait for the developer to fix it.
Future rules may introduce an AI model "passport" with a description of training data, error and "hallucination" control, responsible employees, and requirements for operation within the Russian contour. The use of external platforms may also be restricted, and banks may be obliged to foresee the possibility of changing suppliers in advance. For now, this is just a proposal, not approved requirements.
Read more on the topic:
- State Duma gives start to "sovereign AI": bill passed in first reading
- Neural Networks for Investments: AI Learns to Find Promising Projects
- Nabiullina sets boundaries for AI: neural networks help the Central Bank, but don't decide for it