Back to the MIT repository
5. Human-Computer Interaction2 - Post-deployment

Loss of Control Risks

Risks associated with scenarios in which one or more general-purpose AI systems come to operate outside of anyone's control, with no clear path to regaining control. This includes both passive loss of control (gradual reduction in human oversight) and active loss of control (AI systems actively undermining human control)

Source: MIT AI Risk Repositorymit1449

ENTITY

2 - AI

INTENT

3 - Other

TIMING

2 - Post-deployment

Risk ID

mit1449

Domain lineage

5. Human-Computer Interaction

92 mapped risks

5.2 > Loss of human agency and autonomy

Mitigation strategy

1. Mandate the retention of human agency by prohibiting the deployment of general-purpose AI systems in safety-critical roles requiring autonomous, open-ended goal execution. Concurrently, establish protocols requiring human review and explicit approval for all high-impact, AI-assisted operational decisions to prevent passive loss of oversight. 2. Implement continuous monitoring and advanced behavior analytics across the AI's lifecycle, including training and inference endpoints. This strategy must focus on detecting *model drift*, performance anomalies, and early indicators of self-preserving or intentionally misaligned behavior to enable timely intervention and system rollback. 3. Conduct rigorous and repeated adversarial stress testing, such as comprehensive AI red teaming, throughout the development and post-deployment phases. The objective is to proactively uncover and harden defenses against vulnerabilities that could enable active loss of control, deceptive reasoning, and unintended emergent agentic capabilities.