We introduce a governance framework that shifts the focus of AI safety from model performance to the human context of every interaction.
By utilizing tools like the Cognitive Susceptibility Taxonomy and Contextual Vulnerability Overlays, we argue that an identical AI response can be helpful in a casual setting but dangerous during a personal crisis. We emphasize that situational fragility, such as grief or youth, requires stricter deployment thresholds to prevent the machine from distorting a user’s autonomy or reality.
To mitigate these risks, we advocate for structural safeguards like no-command defaults, higher verification friction, and mandatory human handoffs in high-stakes scenarios. True safety is not an abstract metric but a dynamic response to the specific vulnerabilities of the person using the system.
Article related to the podcast available here
