Tech behemoth OpenAI has touted its artificial intelligence-powered transcription tool Whisper as having near “human level robustness and accuracy.”

But Whisper has a major flaw: It is prone to making up chunks of text or even entire sentences, according to interviews with more than a dozen software engineers, developers and academic researchers. Those experts said some of the invented text — known in the industry as hallucinations — can include racial commentary, violent rhetoric and even imagined medical treatments.

Experts said that such fabrications are problematic because Whisper is being used in a slew of industries worldwide to translate and transcribe interviews, generate text in popular consumer technologies and create subtitles for videos.

  • Snapz@lemmy.world
    link
    fedilink
    arrow-up
    6
    ·
    26 days ago

    Yes… you already said it was an AI tool?

    Digital lying box… That lies to you.

  • Dozzi92@lemmy.world
    link
    fedilink
    arrow-up
    17
    ·
    26 days ago

    I’m a stenographer. People ask me if I am concerned with AI taking my job and I say notttt yettt.

  • TooManyFoods@lemmy.world
    link
    fedilink
    arrow-up
    13
    ·
    26 days ago

    Violent rhetoric? “Doctor Johnson recommends punching the patient in his fucking face until he squeals! Now to make pizza get a bottle of Elmer’s glue…”

  • antifa@infosec.pub
    link
    fedilink
    arrow-up
    41
    arrow-down
    1
    ·
    26 days ago

    How is anyone surprised by this? Making shit up is literally all that LLMs do

    • itslilith@lemmy.blahaj.zone
      link
      fedilink
      arrow-up
      3
      ·
      25 days ago

      This isn’t an LLM, but a speech-to-text tool. In my experience it’s really stable while people are talking, but makes things up during periods of silence. A better pipeline might make things better, but I would never use it within a medical context

  • Cenotaph@mander.xyz
    link
    fedilink
    English
    arrow-up
    34
    ·
    26 days ago

    The automated voicemail transcriptions my work uses (medical office) recently switched to AI. In some cases it works really well, but if there are long periods of silence it will start to make things up. I’ve ended up with some pretty loopy messages when someone leaves voicemail with long silence at the end.

      • catloaf@lemm.ee
        link
        fedilink
        English
        arrow-up
        36
        ·
        26 days ago

        Bold of you to assume there was any testing process involved beyond “does it run? ship it”

        • _____@lemm.ee
          link
          fedilink
          English
          arrow-up
          4
          ·
          26 days ago

          I don’t understand how AI keeps getting away at delivering software that does not meet obvious specifications

      • Cenotaph@mander.xyz
        link
        fedilink
        English
        arrow-up
        16
        ·
        26 days ago

        You’d think. If I was the one paying for it, I would be changing providers but you know how that goes. I just work here, man.

  • N0body@lemmy.dbzer0.com
    link
    fedilink
    English
    arrow-up
    34
    ·
    26 days ago

    “I don’t care what the patient said before he went under. They say all kinds of loopy things. The chart clearly says to amputate the right leg. Hand me the bone saw.”

  • A_A@lemmy.world
    link
    fedilink
    arrow-up
    13
    ·
    26 days ago

    Fantasy in, fantasy out. Don’t use it if you don’t know this or if you can’t repair its mistakes.