For one month beginning on October 5, I ran an experiment: Every day, I asked ChatGPT 5 (more precisely, its “Extended Thinking” version) to find an error in “Today’s featured article”. In 28 of these 31 featured articles (90%), ChatGPT identified what I considered a valid error, often several. I have so far corrected 35 such errors.

  • OpenStars@piefed.social
    link
    fedilink
    English
    arrow-up
    2
    arrow-down
    1
    ·
    4 hours ago

    Using ChatGPT to “fix” Wikipedia, what could possibly go wrong? (/s as the approach seems valid, this is just a funny statement)

  • Brave Little Hitachi Wand@feddit.uk
    link
    fedilink
    English
    arrow-up
    16
    ·
    9 hours ago

    Speaking very generally, it’s still conceding an amount of human intelligence and there are problems with it that are worth talking about, but it’s a use of AI that at least defers to human judgment, and as long as users are still personally researching and writing their own edits I honestly don’t hate it. Much.

    • lime!@feddit.nu
      link
      fedilink
      English
      arrow-up
      11
      ·
      8 hours ago

      it’s mostly outsourcing attention, which is pretty acceptable for a large project like wikipedia.

      • Bldck@beehaw.org
        link
        fedilink
        English
        arrow-up
        1
        ·
        2 hours ago

        That’s my main use for LLMs

        • I write the code logic, the main argument points, etc
        • let the LLM lint, format and structure the discussion
        • I provide another round of copy editing, styling and other updates
        • lime!@feddit.nu
          link
          fedilink
          English
          arrow-up
          1
          ·
          2 hours ago

          personally i have separate linters, formatters and structure markers that don’t raise the temperature of my apartment when in use, but you do you.

      • Brave Little Hitachi Wand@feddit.uk
        link
        fedilink
        English
        arrow-up
        9
        ·
        edit-2
        6 hours ago

        Right - I won’t call it a good thing to let people de-skill on reading comprehension skills, but they’re donating their labour to a public benefit! I’m hardly going to scold them as if I was their professor.

        • lime!@feddit.nu
          link
          fedilink
          English
          arrow-up
          7
          ·
          7 hours ago

          my thought is mainly that there aren’t enough hours in the day to read and check everything on wikipedia. there’s a reason the scots vandalism went unnoticed so long, people just don’t have the time.

  • Qwel@sopuli.xyz
    link
    fedilink
    English
    arrow-up
    8
    ·
    edit-2
    8 hours ago

    Most of the errors aren’t so bad, but it’s definitely nice to correct them.


    You need to know Wikipedia’s system a bit though, because ChatGPT suggests these kind of things:

    Want me to draft a crisp correction note you can paste on the article’s talk page?

    Using LLMs when interacting with other editors is “strongly frowned upon”, and you can get banned if you refuse to stop. Especially if you are editing a lot of pages as you just discovered a lot of issues.