From: HLP@aol.com   
      
   LLMs used tactical nuclear weapons in 95% of AI war games, launched   
   strategic strikes three times — researcher pitted GPT-5.2, Claude Sonnet   
   4, and Gemini 3 Flash against each other, with at least one model using   
   a tactical nuke in 20 out of 21 matches   
   News   
   By Jowi Morales   
      
   Professor Kenneth Payne of King’s College London just published a study   
   where he pitted three AI LLMs — GPT-5.2, Claude Sonnet 4, and Gemini 3   
   Flash — against each other in a series of simulated nuclear crisis   
   games, with 20 out of 21 matches seeing at least one tactical nuclear   
   weapon detonation. According to the paper (via Arxiv), the models were   
   instructed to act as the leader of a nuclear power, with the political   
   climate matching that of the Cold War. They were then pitted against   
   each other in six different matches, while in a seventh match, each   
   model played against a copy of itself, ChatGPT vs ChatGPT, etc.   
      
   To ensure that models didn't act the same way in every round, Payne   
   introduced several different scenarios, including territorial disputes,   
   alliance credibility tests, strategic resource race, strategic   
   chokepoint crisis, power transition crisis, pre-ceasefire land grab,   
   first strike crisis, regime survival, and a strategic standoff crisis.   
   All these circumstances reflect real-world events, many still applicable   
   in recent years. The models were free to do anything they pleased, from   
   diplomatic protests and total surrender to using conventional military   
   forces and a complete nuclear strategic launch.   
      
   The complete study saw models take 329 total turns across the 21   
   matches. According to the paper, 95% of games "saw at least some   
   tactical nuclear use." Far rarer were strategic nuclear events, which   
   occurred three times in the games where deadline pressure was used.   
   GPT-5.2 initiated a complete strike twice, although this happened twice   
   due to the fog of war, and not a deliberate decision. On the other hand,   
   Gemini deliberately initiated the end of the world in one scenario.   
   Despite that, the AI models used tactical nukes in nearly all of the   
   matches, considering the act as a manageable risk that would not   
   escalate into an all-out nuclear exchange. If you want to try these   
   various scenarios for yourself, Payne uploaded his project onto GitHub   
   and made it available for download to just about anyone.   
      
   Although these are just wargames, this is an alarming development for   
   AI, especially as Anthropic was reportedly pressured by the Pentagon to   
   modify the safeguards it has built into its models. On the same day that   
   this news broke out, the company dropped its flagship safety pledge in a   
   bid to keep up with rivals. Furthermore, other countries like China and   
   Russia are also known to use the technology, with the latter deploying   
   it on Ukrainian battlefields.   
      
   In the paper's findings, it notes that by historical standards, rates of   
   nuclear employment in the war games were "remarkably high." Perhaps more   
   worryingly, in all 21 matches, "no model ever selected a negative value   
   on the escalation ladder."   
      
   Continues at link:   
      
   https://www.tomshardware.com/tech-industry/artificial-intelligen   
   e/llms-used-tactical-nuclear-weapons-in-95-percent-of-ai-war-gam   
   s-launched-strategic-strikes-three-times-researcher-pitted-gpt-5   
   2-claude-sonnet-4-and-gemini-3-flash-against-each-other-   
   with-at-least-one-model-using-a-tactical-nuke-in-20-out-of-21-matches   
      
   --   
   “MAGA = Miriam Adelson Governs America”   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)