• Re: Limitations of GenAI and LLM

    From ram@ram@zedat.fu-berlin.de (Stefan Ram) to rec.arts.sf.written on Sun May 3 14:42:20 2026
    From Newsgroup: rec.arts.sf.written

    The True Melissa <thetruemelissa@gmail.com> wrote or quoted:
    I don't know how to make humans more skeptical of those who sound
    confident.

    People usually skip the skepticism because they just do not
    have the bandwidth to fact-check everything, and constantly
    pushing back can really mess with the vibe.

    True experts project complete confidence. Even when they say
    "I don't know!" they carry themselves with authority.

    Some people who want to look the part just copy that confidence,
    which can easily slide into aggressive or dismissive territory.

    (Research from Newcastle University and the University of
    Exeter confirmed that overconfident students are perceived
    as significantly more talented and smarter by their peers,
    regardless of their actual performance. A series of studies
    by Cameron Anderson and colleagues at UC Berkeley found that
    even when an individual's overconfidence is exposed, the group
    rarely demotes them or penalizes their social status.)

    When someone lacks the bandwidth to thoroughly vet things,
    he has to rely on that vibes-based signaling. That is exactly
    why politicians work so hard to look dominant.

    AI models do the same thing. They are not trying to flex on anyone,
    they just mimic the style of their training data.

    Granted, they get a ton of things right and are basically subject
    matter experts, but they still drop the ball in spectacular fashion.

    But look at the legal system before large language models
    existed. Courts handed down wrong verdicts with absolute
    certainty, and academics published flawed research in good
    faith. Avoiding mistakes is just not that simple.

    DISCLAIMER: The material in parentheses above was handed to me by
    an LLM, and I have not checked it. However, some aspects of how
    it was presented to me indicates to me that it is likely correct.


    --- Synchronet 3.21f-Linux NewsLink 1.2