Automated image generators are often accused of spreading harmful stereotypes, but studies usually only look at MidJourney. Other tools make serious efforts to increase diversity in their output, but effective remedies remain elusive.

  • Terces@lemmy.world
    link
    fedilink
    arrow-up
    28
    arrow-down
    1
    ·
    1 year ago

    I wonder if this is because AI is trained on data that ‘is’ and has therefore no concept of how it ‘should be’. Maybe it is an effective mirror of society…

    • Actually it is trained on data ‘the trainers have’. This is different from ‘trained on data that “is”’ or any other idealized view of data.

      Data that ‘the trainers have’ is always an incomplete view of anything, and adding meaningfully to datasets is always very difficult.

      • Terces@lemmy.world
        link
        fedilink
        arrow-up
        2
        ·
        1 year ago

        I may have oversimplified my statement. Of course an objective description of reality is impossible. A curse on all social sciences and statistics.

        My post was more a showerthought…even if the data is incomplete, whatever THAT data implies will also be the stereotype the AI will learn. Misrepresentation of minorities in sample data is absolutely nothing new. But even if the data WAS complete, it would probably still be very biased. I think we often don’t notice structural discrimination and AI would simply reproduce those and confront us with it. In that sense I think it is a very interesting way to get a sort of ‘outside look’ at our own society and that is something that’s very useful.

    • ubermeisters@lemmy.world
      link
      fedilink
      arrow-up
      16
      arrow-down
      1
      ·
      1 year ago

      That’s 100% a real issue. Fortunately for all these clickbait articles, most people don’t really grasp how these things are trained or how input data affects them during training.

      • Car@lemmy.dbzer0.com
        link
        fedilink
        arrow-up
        4
        ·
        edit-2
        1 year ago

        And even if we could provide the training algorithm a perfectly diverse dataset, who gets to decide what that means? You could probably poll a million anthropologists from across the world and observe trends, but no certain consensus. What if polling anthropologists in underdeveloped nations skews in a different direction than what we consider rich countries? How about if a country was a colonizer in the past or has participated in a violent revolution?

        How do we decide who qualifies as an anthropologist? Is a doctorate required, or is a college degree with numerous publications sufficient?

        I don’t think we’ll ever see a perfectly neutral solution to this problem. At best, we can come equipped with knowledge that these tools may come with some biases, like when you analyze texts from the past. You make the best with what you have and strive to improve