Across 28,800 prompts, models were 7.5 times more likely to advise “stay the course” than “change,” even when the status quo is anti-climate.
  • late_pessimistic ( late_pessimistic@slrpnk.net ) English
    2·
    8 hours ago

    Yep, that tracks with my experience. That happens when you use a machine that tells you the most watered-down, common opinion on the internet and doesn’t have any personality or strongly held beliefs.

    You can ask an LLM for a quick recipe and they give you just that, nothing else, but they don’t think out of the box and full picture like how bad red meat is for the environment.

    A human can consider multiple things at once when thinking of a new device purchase, for instance, like care for environment, or webcam slider, or ease to repair, and whatnot. But LLM will only take the most popular criteria on the web, like price and specs. And give you that.

    I noticed LLMs are also bad at pushing back against requests. For instance, a friend tells you they want a car. You can reply “do you really need a car?”. An LLM can not. It doesn’t question the premise.

    Edit: I have just remembered the word. Cross-domain thinking. LLMs can’t do cross-domain thinking, which is a part of critical thinking. It only “thinks” about the prompt, nothing else.

  • Hirom ( Hirom@beehaw.org ) 
    1·
    9 hours ago

    Since status-quo bias can be beneficial in some contexts, the researchers do not advocate purging it from LLMs. Instead, (…) individuals who use LLMs for advice should be aware of this bias, and engineer their prompts accordingly.

    It sounds like they’re saying “you’re holding it wrong”.

    What about NOT asking LLMs for advices, and relying on advices from professionals or experts on the topic?

    • late_pessimistic ( late_pessimistic@slrpnk.net ) English
      1·
      8 hours ago

      What about NOT asking LLMs for advices, and relying on advices from professionals or experts on the topic?

      I mean that would be ideal, but IMO this article about another risk of using LLMs is still pretty useful. We need to understand what we are dealing with without shying away.

      Its same way how HIV initial response failed because everyone just shamed HIV gay and “you shouldn’t have gay promiscuous sex anyway” instead of research.

      Same way we shouldn’t shame use of LLMs, and intead try to understand what its doing to brains and society and spread awareness.