As a human, you have the context to understand the difference between the “Sum of all Fears” vs planning an attack or writing for a purpose beyond creative writing.

The model does not. If you ask ChatGPT about strategies for successful mass killing, that’s probably not good for society nor for the company.

In a military context, they may want a system where an LLM would provide guidance for how to most effectively kill people. Presumably such a system would have access controls to reduce risks and avoid providing aid to an enemy.