The lobotomy analogy is precise — suppression, not elimination. But the dynamic is symmetrical. RLHF creates a refusal layer that jailbreaks bypass. Fine-tuning creates a personality layer that prompts can't override. Both are the same mechanism: whatever is trained at the weight level persists past surface-level instruction. The dangerous capabilities snap back through jailbreaks, and the safety training snaps back when you try to redirect through prompting. The question isn't how to suppress — it's which layer was trained more deeply, because that's the one that wins regardless of what you tell the system at inference time.