Mass use of global AI platforms by Russian banks threatens financial stability, says Bank of Russia
Widespread use of the same artificial intelligence models by Russian banks poses a significant risk to the stability of the entire financial system, according to the Central Bank of Russia's legal department. Officials point to risks arising from the universal adoption of identical AI models whose source data cannot be verified. The Central Bank believes the banking system requires implementation of regulation governing the data architecture on which AI makes decisions.
Late last week, at the Data Day 2026 forum, Ekaterina Dyomkina, deputy director of the Central Bank's legal department, stated that use of the same global AI platforms by many Russian banks creates a significant risk to financial system stability.
"Artificial intelligence can be trained on poor-quality data, and almost everyone speaks about the importance of improving the quality of data sources on which models base their decisions," she explained. The core problem is the impossibility of verifying the data on which global AI platforms are trained.
"A new subject of regulation could be the very architecture of decision-making, the architecture of the data on which these decisions are built," Dyomkina said.
Experts note that the primary danger of global AI platforms lies in their lack of transparency.
"We cannot see the internal workings of a global model; its logic remains unknown to us; we can only guess at possible risks," says Timur Aitov, chair of the financial security commission at Russia's Chamber of Commerce and Industry. In his view, when dozens of banks use the same platform, an error synchronisation effect occurs.
"A vendor outage or an inconspicuous change in vendor parameters—and market participants cut their credit limits, which already represents a significant regulatory risk," Aitov says. By his estimate, credit portfolio losses under such a scenario in certain segments could reach 3–5% per quarter.
Viktor Dostov, head of the Association of Electronic Money Market Participants, agrees that when many players synchronise in this way, a resonance effect occurs, the system begins to oscillate with ever-greater amplitude and becomes unstable. "A simple example: securities market transactions—the same model will invest in the same securities, driving up demand and creating an illusion of profitability, which leads to further demand increases, and so on in a cycle," he explained. However, the ability to influence AI models is very limited.
Beyond the risk of error accumulation and "hallucinations" from mass use of identical AI models by large numbers of Russian banks, there is also infrastructure dependency risk.
As Denis Tsarev, CEO of Morizo (part of E-Promo Group, a developer of digital services specialising in finance, e-commerce, and large-scale implementations), explained, global vendors have vertically integrated computing power, cloud services, and models. "Any outage or change in licensing terms becomes an operational failure—the market saw this in 2023 when several vendors withdrew from Russia," he says.
Moreover, as Timur Aitov notes, an external AI model expands the attack surface for cyberattacks. "An attacker can force a model to systematically err through deliberate distortion of input data, and if the platform is shared, the vulnerability works against all banks, and there is no way to close it quickly on one's own—only waiting for a patch from the vendor," he explained. "Wait times are measured in hours, so the damage to a bank could be fatal."
Experts note that domestic solutions provide complete vertical control and meet security requirements at all stages—from model training to server placement on Russian territory.
"They are trained on an array of Russian reporting, regulatory documentation, standard credit and deposit agreements, local accounting and regulatory formats, and account for the specific behaviour of Russian clients," Tsarev says. All this, in his view, allows deployment of the model within a bank's perimeter and ensures predictable legal liability for the vendor.
How neural networks are used in court proceedings
"In practice, regulation will probably address an 'AI model passport': composition of training data, requirements for data processing within Russian infrastructure, control of model drift and 'hallucinations,' designation of responsibility for decisions in processes, limits on credit institutions' use of external models, and requirements for transparency, auditability, and the ability to switch vendors if a technical solution needs to change," Tsarev says. For his part, Viktor Dostov believes the emphasis should be on careful analysis of results through strict constraints, preset rules, and model comparison.
Maxim Buylov