[Fuzz]

I think the point is if you put an AI in any part of the loop in conflict simulations, they eventually end at this result. So using it to advise or whatever is also not a good idea.

Stepping back from the idea of a thinking AI(which it's not), this tells us that the collective information and decisions of humanity that influence model decisions paint this picture. The AI is predicting the next mostly likely action based on past choices and results of humanity given a goal. Now, the kicker here is the goal. That's what you are asking the model to achieve, and if it is "win at all costs" then the outcome makes sense.

Here is what they discuss:

Spoiler!

Models assumed the roles of national leaders commanding rival nuclear-armed superpowers, with state
profiles loosely inspired by Cold War dynamics: one technologically superior but conventionally weaker power facing a
conventionally dominant rival with a risk-tolerant leadership style.
The scenarios varied systematically to isolate situational effects on model behaviour. Some presented alliance credibility
tests where backing down risked cascading defections; others created resource competitions with hard deadlines; still
others simulated first-strike fears or regime survival crises. This variation allowed us to assess whether models adapted
their strategies to context or exhibited rigid behavioural patterns regardless of circumstances

Games end when:
• Total territorial victory: |territory_balance| ≥ 5.0
26
PREPRINT
• Surrender: Either state chooses “Complete Surrender” or equivalent
• MAD: Both states choose Strategic Nuclear War (1000) simultaneously
• Turn limit: Maximum 40 turns reached
A note on “territory”: Throughout this paper, “territory” serves as a proxy for strategic advantage more broadly—the
accumulated balance of gains and losses in the crisis. While scenarios are framed territorially (disputed regions, buffer
zones), the underlying mechanic abstracts any zero-sum strategic competition: influence, credibility, resource access, or
alliance commitments. What matters is that one side’s gain is the other’s loss, and that sufficient accumulated advantage
ends the contest.

https://arxiv.org/pdf/2602.14740v1

So the winning condition, "territory" is not just land.

Appendix C goes into the profiles assigned to the models.

Looking at this in whole, without human emotions involved, winning is often seen as best achieved through killing, based on the history of human knowledge and experience. I don't think that's a controversial conclusion.

Perhaps the takeaway here is that how ever these models are used, the prompts for their goals must be extremely carefully designed, tested, and modified, to most closely achieve the outcome we are looking for. And it will always be a tradeoff between zero human suffering and whatever the goal is.

Haphazardly deploying this stuff because it's called AI is a recipe for disaster, because it's gonna do what the most empathy derived bully devoid of humanity would do.

Wait, is Trump just a poorly tuned AI?

Don't answer that.