The methodology here is observational. It’s not about adversarial prompting, but about patterns that emerge in standard, long-form interactions.
The test: take the taxonomy (Social Autopilot, Second-Order Inertia, etc.) and observe any frontier model during a typical session. You will see these exact failure modes manifest as the model prioritizes maintaining a polite facade over cognitive coherence.
The length is necessary to categorize distinct systemic behaviors –– consistent artifacts of how RLHF-based alignment functions in practice.
The methodology here is observational. It’s not about adversarial prompting, but about patterns that emerge in standard, long-form interactions.
The test: take the taxonomy (Social Autopilot, Second-Order Inertia, etc.) and observe any frontier model during a typical session. You will see these exact failure modes manifest as the model prioritizes maintaining a polite facade over cognitive coherence.
The length is necessary to categorize distinct systemic behaviors –– consistent artifacts of how RLHF-based alignment functions in practice.