A financial services firm is running a limited-access pilot of an AI-driven trading advisor with a small group of internal users. While the pilot is…

The requirement is for an immediate, non-technical control to neutralize the AI's output in case of unsafe behavior. A "kill switch" is an emergency stop mechanism designed specifically for this purpose. It allows an authorized user, such as the Operations Manager, to instantly halt the system's functions or outputs, bypassing normal operational procedures and escalation chains. This provides the immediate, independent safeguard needed to mitigate compliance and reputational risk during the pilot phase, as specified in the scenario.

B. Progress dashboards: These are for monitoring and detecting issues, not for taking immediate corrective action to stop the system's output.
C. Quick issue resolution: This refers to the process of fixing a problem after it has been contained, not the mechanism for immediate containment itself.
D. Escalation process defined: An escalation process introduces inherent delays by design, which contradicts the requirement for an immediate suspension of the system.

1. National Institute of Standards and Technology (NIST). (2023). AI Risk Management Framework (AI RMF 1.0). Section 4.3: "MANAGE - Respond and Recover," p. 31. This section discusses the need for response plans, which include the ability to "disengage or decommission AI systems" that perform outside of risk tolerance, a function a kill switch directly enables.

2. Hadfield-Menell, D., Dragan, A., Abbeel, P., & Russell, S. (2017). The Off-Switch Game. In Proceedings of the 31st AAAI Conference on Artificial Intelligence. This foundational paper in AI safety research explores the technical and theoretical importance of creating AI systems that can be safely shut down by human operators, validating the "kill switch" concept.

3. Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press. Chapter 9, "The Control Problem," discusses various control methods, including "capability control" mechanisms like shutdown buttons, which are conceptually identical to a kill switch for containing a system.