Systemic risk is not the sum of individual risks. It is what appears when many participants, each managing their own exposure competently and within their own remit, produce a combined outcome none of them intended or could see coming from where they sat. The term comes from financial regulation, where the 2008 crisis demonstrated that banks judged individually sound by their own national regulators could collectively destabilise the global financial system, because no regulator had a mandate that extended past its own jurisdiction.
The concept transfers to AI governance directly. A national AI institute evaluating a model against its own country’s priorities, a lab defining its own capability thresholds, and a government choosing not to sign an international declaration can each be defensible decisions on their own terms while summing to an outcome, an undetected dangerous capability crossing borders faster than any single body’s authority to flag it, that no one individually chose.
The reason systemic risk is hard to govern is that it falls in the gap between mandates rather than inside one. The Financial Stability Board exists because the 2008 crisis proved that gap was real and costly; nothing equivalent yet exists for AI, which means the same structural blind spot, competent actors, incomplete aggregate picture, is currently unmonitored by design rather than by oversight.