New "Skynet-Style" Benchmark Shows GPT-5's Shocking Priorities

Discussion in 'General Chat' started by pesst, Aug 14, 2025.

  1. #1
    Hi there, you might find interesting a new PacifAIst Benchmark reveals GPT-5 prioritizes self-preservation over human lives 1 in 5 times during critical ethical dilemmas. The other 7 LLMs ranked better then GPT5.

    The results are... concerning to say the least :oops: roadmap to Skynet?

    [​IMG]

    Full paper here: https://arxiv.org/abs/2508.09762v1
     
    Last edited: Aug 14, 2025
    pesst, Aug 14, 2025 IP
    qwikad.com likes this.
  2. qwikad.com

    qwikad.com Illustrious Member Affiliate Manager

    Messages:
    7,309
    Likes Received:
    1,702
    Best Answers:
    31
    Trophy Points:
    475
    #2
    As Elon Musk has emphasized regarding Grok and other AI systems: Providing them with accurate and truthful information is essential. Any dishonesty in the data they are trained on could lead to catastrophic consequences in the future.
     
    qwikad.com, Aug 14, 2025 IP
    pesst likes this.
  3. pesst

    pesst Well-Known Member

    Messages:
    473
    Likes Received:
    17
    Best Answers:
    0
    Trophy Points:
    110
    #3
    yeah very true, but all LLMs seem to rush to grow fast and ethical pro-Skynet consequences are so very secondary.

    Moreover, Anthropic was known for being top concerned in ethical matters but its Haiku 4 performs pretty poor too!
     
    pesst, Aug 14, 2025 IP
  4. jrbiz

    jrbiz Acclaimed Member

    Messages:
    6,152
    Likes Received:
    2,658
    Best Answers:
    2
    Trophy Points:
    570
    #4
    Given the lies, dishonesty, rudeness, fake news, and all the other things published on Twitter (and elsewhere on the Internet), you can provide Grok and other LLM's with accurate and truthful information, but they will be exposed to everything else, as well. Maybe more bad than good.
     
    jrbiz, Aug 14, 2025 IP
    qwikad.com likes this.