Systemic Risk Is Now Cognitive
In 2024 the Financial Stability Board published a report on systemic risk associated with the adoption of artificial intelligence in financial markets. The central concern was not volatility, liquidity, or credit, but concentration. Three or four providers of foundational models control the cognitive infrastructure used by most global financial institutions. A failure, vulnerability, or strategic decision by any one of these providers can propagate simultaneously across the entire system.
The FSB had examined concentration risk in cloud services in 2023, identifying Amazon Web Services, Microsoft Azure, and Google Cloud as critical nodes for the operational resilience of the financial sector. With foundational models the issue reappears at a higher level. Dependence has shifted from the infrastructure that runs systems to the infrastructure that produces the analyses, recommendations, and risk assessments on which systems base their decisions.
This is a qualitatively different form of systemic risk. Liquidity risk propagates through balance sheets. Operational risk propagates through infrastructures. Cognitive risk propagates through shared models of interpretation. When institutions use the same models to evaluate the same risks, the system loses the interpretive diversity that historically allowed some actors to stand on the other side of the market during crises and absorb shocks.
Haldane (2011) described this dynamic in the context of algorithmic trading: the complexity of the system had exceeded the cognitive capacity of those responsible for governing it. With foundational models the structure is identical but the scale is larger. Trading algorithms acted on markets. Foundational models act on the reasoning that precedes and orients market decisions. The level of abstraction is higher and the visibility of risk correspondingly lower.
The regulatory challenge is substantial. How does one regulate a risk that is not located in a single institution, is not measurable with traditional tools of financial risk, and does not manifest as direct loss until the system is already in crisis? Regulators are applying conceptual frameworks designed for market risk to a risk that has epistemic nature; the tools are not suited to the task. The distance between the problem and the available instruments is the system's primary vulnerability today.
Cognitive systemic risk is not a future possibility. It is a present condition of the global financial system. The question is whether those responsible for governing that system will recognise it before the system demonstrates it.