(cont’d)

…with “key takeaways” and regurgitated paragraphs that all follow the same format. It’s gross, and yet it generates an article long enough with enough keywords to show up on Google.

  • fearout@kbin.social
    link
    fedilink
    arrow-up
    20
    ·
    edit-2
    1 year ago

    Reposting my comment from another similar thread to show that this is easily fixable, and you should be wary of any non-reputable news source anyway.

    So I was curious how current LLMs might handle this with proper instructions, so I asked chatGPT this: “What can you tell me about this Reddit post? Would you write a news article about this? Analyze the trustworthiness of this information:” and pasted the text from the post. Here’s a part of its reply:

    This Reddit post appears to be discussing updates in the context of […] Hearthstone and World of Warcraft. However, there are several factual inaccuracies which raises questions about the trustworthiness of the post:

    1. Hearthstone was first released in 2014, not 1994. This error may be a typo or a misunderstanding by the user.
    1. Glorbo does not appear to be a known character in either the Hearthstone or World of Warcraft universe […]
    1. Chen Stormstout is a known character from the Warcraft universe, but the claim about him being the end boss of a new version of Karazhan doesn’t match with any known information […]
    1. Dragonflight is not known as a video game developer or associated term in the context of Hearthstone or World of Warcraft […]

    Given these points, it seems this post might be either a joke, speculation, or misinformation. […] So I’d recommend taking this information with a grain of caution and verify it from reliable sources.

    […] I would caution against using the information in this post for a news article due to the factual inaccuracies and lack of verifiable sources. Good journalism relies on accuracy, fact-checking, and use of reliable sources. This Reddit post contains several claims that contradict established knowledge, and without further verification from reliable, official sources (like game developers or official news releases), it would not be advisable to use this information in a news article.

    So it’s not even an issue with current models, just bad setup. An autoGPT with several fact-checking questions added in can easily filter this stuff.

      • fearout@kbin.social
        link
        fedilink
        arrow-up
        4
        ·
        edit-2
        1 year ago

        Half of the deleted […] things are chatGPT mentioning its 2021 knowledge cutoff and suggesting double-checking that info. It was mentioned in this case as well.

        If it were an autoGPT with internet access, I think these would prompt an automated online lookup to fact-check it.