4 ms·

As someone who works in CG and has a strong interest in AI for almost 40 years, I have to say the examples you gave give me the opposite impression. I think the

• sgt101 5y ago
  Oh sure the technical outcome is great, but this kind of output (ie. hugely gender biased) is extremely problematic. Saying "oh the input screwed it" is very very naughty. This kind of tool will be used. If it's used it will do damage becuase it's damaging as it is.
  • oh_sigh 5y ago
    If you ask for a nurse as an example, and 85% of nurses are female, shouldn't you expect it to give you a female?
    • 6gvONxR4sf7o 5y ago
      If you ask for a nurse and 85% are female, but 95% of the nurse images online are female, what distribution do you want it to return? “What do you expect it to do” is a worse question than “What do you want it to do?” It’s a problem is that right now there’s only one option (reflect the training set with a bias towards the most common cases).
    • karpierz 5y ago
      It depends on whether you want it to show what someone would expect a nurse to look like, or show what a nurse. Nothing in the definition of a nurse implies female. If you ask for a female US president, should the application simply return a black screen? If you ask for a black US president, would you expect it to only ever return pictures of Barack Obama?
      • Rerarom 5y ago
        Well, nurse ends in an e, which is usually a feminine ending.
        • na85 5y ago
          Not really in English. Uncle, Prince, Sire, and Duke come immediately to mind. Most of their feminine cognates end in consonants.
      • oh_sigh 5y ago
        I agree nothing in the definition of nurse implies female, but something in our culture definitely has "nurse -> female". Likewise, nothing in "college student" definitionally implies age, but I wouldn't consider it a success if my system displayed a picture of a 9 year old.
      • lovemenot 5y ago
        >> Nothing in the definition of a nurse implies female. Unfortunately it's not a great example with which to make this point. Before Florence Nightingale, a nurse was a person who was employed to suckle your babies. Of course 100% were female. Wet nurse was a later coining to differentiate the two meanings.
  • jgerrish 5y ago
    Like, this is amazing. As an accomplishment, very cool, I'd be proud. But you're right, is somebody's finger or faulty data on the scales? Because that's how you get a bumper bowling world laser light show clown world instead of life. The metaphor isn't perfect, and cuts several ways, but it's what my mind came up with.
    • freeone3000 5y ago
      One issue is the data is accurate to the world it is trained on, but not the world we want to exist in. The most likely nurse is female, so this generates only female nurses. The most likely CEO is male, so this generates only male CEOs. And thus we see in-built biases amplified.
  • kingcharles 5y ago
    I absolutely agree the output is problematic without a horrible disclaimer. We've seen what happens when this technology is used to make financial and employment decisions, and in the justice system the results can be tragic. This demo toy just pushes the bias right in your face and makes the invisible visible, which could be very useful for highlighting the problem for lawmakers.
  • planetsprite 5y ago
    The gender bias exists due to the assumptions it makes on the sample data. I'm sure if you searched "<gender> <profession>" it would get you what you want. If 95% of publicly available images of a profession are of one gender, should the tool be deliberately modified so that it's X% instead? I totally agree that models that purport to represent the world can be hugely biased and reconfirming of negative biases, but how would avoiding representing those biases be achieved? Strict gender ratios in training data?
    • sdenton4 5y ago
      There are ways to improve the situation. Here's one way... Given a text description, sample from the space of /more restrictive/ text descriptions (eg, add adjectives) and then draw pictures for those. Then modify the sampling space over more descriptive sentences to equalize on gender or other axes.
    • makeitdouble 5y ago
      It’s more complicated this specific tool itself IMHO. Stock photos have the same problem, where asking for a profession or a category will give mostly one kind of representation, and it’s up to the user to go dig further to find different representations. DALL-E being on par with stock photos biases could seem benign, but it also means these issues get propagated further more down the line, and the more AI generated images get popular, the worse it gets cemented (“it has always been that way”) and could actually displace niches where better images were being used until then.