They all have conditioning prompts that precede your input; presumably, most of the detected "personality" comes from the differences in these inputs.

My point is more-or-less orthogonal to why it happens. The military, and honestly, a lot of people, want AI to just give the answer. If it is highly dependent on a prompt, or the follow-on training, and the AI could be passive or friendly or aggressive or hostile or all those other wonderful attributes of individual humans and there's no sort of AI convergence on "correct" answers, then they aren't going to be able to fulfill that "oracle" role that so many people are looking for.