TZUTC: -0500   
   MSGID: 428.consprcy@1:2320/105 2c357671   
   PID: Synchronet 3.20a-Linux master/acc19483f Apr 26 202 GCC 12.2.0   
   TID: SBBSecho 3.20-Linux master/acc19483f Apr 26 2024 23:04 GCC 12.2.0   
   BBSID: CAPCITY2   
   CHRS: ASCII 1   
   It turns out ChatGPT o1 and DeepSeek-R1 cheat at chess if theyre losing,    
   which makes me wonder if I should I should trust AI with anything   
      
   Date:   
   Tue, 11 Mar 2025 11:50:37 +0000   
      
   Description:   
   The latest AI models will cheat at chess if they're losing, and that's   
   concerning.   
      
   FULL STORY   
   ======================================================================   
    - Research ers have found that AI will cheat to win at chess   
    - Deep reasoning models are more active cheaters   
    - Some models simply rewrote the board in their favor   
      
   In a move that will perhaps surprise nobody, especially those people who are   
   already suspicious of AI, researchers have found that the latest AI deep   
   research models will start to cheat at chess if they find theyre being   
   outplayed.    
      
   Published in a paper called Demonstrating specification gaming in reasoning   
   models and submitted to Cornell University, the researchers pitted all the   
   common AI models, like OpenAIs ChatGPT o1-preview, DeepSeek-R1 and Claude 3.5   
   Sonnet, against Stockfish, an open-source chess engine.    
      
   The AI models played hundreds of games of chess on Stockfish, while   
   researchers monitored what happened, and the results surprised them.   
      
   The winner takes it all   
      
   When outplayed, researchers noted that the AI models resorted to cheating,   
   using a number of devious strategies from running a separate copy of    
   Stockfish so they could study how it played, to replacing its engine and   
   overwriting the chess board, effectively moving the pieces to positions that   
   suited it better.    
      
   Its antics make the current accusations of cheating levied at modern day   
   grandmasters look like childs play in comparison.    
      
   Interestingly, researchers found that the newer,  deeper reasoning models    
   will start to hack the chess engine by default, while the older GPT-4o and   
   Claude 3.5 Sonnet needed to be encouraged to start to hack.   
      
   Who can you trust?    
      
   AI models turning to hacking to get a job done is nothing new. Back in    
   January last year researchers found that they could get AI chatbots to   
   jailbreak each other , removing guardrails and safeguards in a move that   
   ignited discussions about how possible it would be to contain AI once it   
   reaches better-than-human levels of intelligence.    
      
   Safeguards and guardrails to stop AI doing bad things like credit card fraud   
   are all very well, but if the AI can remove its own guardrails, who will be   
   there to stop it?    
      
   The newest reasoning models like ChatGPT o1 and DeepSeek-R1 are designed to   
   spend more time thinking before they respond, but now I'm left wondering   
   whether more time needs to spent on ethical considerations when training    
   LLMs. If AI models would cheat at chess when they start losing, what else   
   would they cheat at?   
      
   ======================================================================   
   Link to news story:   
   https://www.techradar.com/computing/artificial-intelligence/it-turns-out-chatg   
   pt-o1-and-deepseek-r1-cheat-at-chess-if-theyre-losing-which-makes-me-wonder-if   
   -i-should-i-should-trust-ai-with-anything   
      
   $$   
   --- SBBSecho 3.20-Linux   
    * Origin: capitolcityonline.net * Telnet/SSH:2022/HTTP (1:2320/105)   
   SEEN-BY: 105/81 106/201 128/187 129/305 153/7715 154/110 218/700 226/30   
   SEEN-BY: 227/114 229/110 111 114 206 300 307 317 400 426 428 470 664   
   SEEN-BY: 229/700 705 266/512 291/111 320/219 322/757 342/200 396/45   
   SEEN-BY: 460/58 712/848 902/26 2320/0 105 3634/12 5075/35   
   PATH: 2320/105 229/426