4 ms·

Prompting, realized, and unrealized bias in generative AI

14 points by andy99 3y ago

jxf 3y ago
Unrelated, but this is the first time I've seen the *.onl TLD.
zarathustreal 3y ago
Bias is a function of perspective. Every entity capable of generating output will be doing so based on "knowledge of" some set of inputs, even if that "knowledge" is part of its "hardware." It seems people are still wasting energy discussing "bias" in generative AI, so hopefully this ends the discussion as such. Obviously it is not possible to generate something without starting from copying something, this is the function of memories. In a physical system, this is t=n-1. Therefore it is not possible for a generative AI model to be free of bias. These discussions always seem to carry a moral connotation by using the word "bias" and relating it to hot-topic biases like racial and gender bias but the reality is that there's nothing necessarily immoral about a generative AI model's biases. Likewise, there's nothing necessarily moral about "fixing" these biases, because the reality is that "fixing bias" in a generative AI model is nothing more than realigning its bias to be more in line with the operator's own biases. This is an important distinction to realize because it helps to put the priority of the work in perspective. With a more accurate perspective of the priority we're better able to identify inefficient use of societal resources, such as taxpayer funding, postdoc researcher hours, and hours of general discourse.
- giraffe_lady 3y ago
  It seems like you're saying it will always have a bias, and one bias is as good as another. But while there may be nothing necessarily immoral about bias existing in general, but there certainly can be something immoral about applying a bias in particular. People aren't specifically focused on the ones you dismiss as "hot button" because of their popularity, they are hot button because those biases powerfully shape the world we live in. It is immoral to reproduce them, and it is morally necessary to choose not to.
  - djyaz1200 3y ago
    [flagged]
    - giraffe_lady 3y ago
      [flagged]
      - dang 3y ago
        Since we've asked you many times to stop breaking the site guidelines and you've continued repeatedly to do it (e.g. here and https://news.ycombinator.com/item?id=37203245 https://news.ycombinator.com/item?id=37203245), I've banned this account. If you don't want to be banned, you're welcome to email hn@ycombinator.com and give us reason to believe that you'll follow the rules in the future. They're here: https://news.ycombinator.com/newsguidelines.html https://news.ycombinator.com/newsguidelines.html. Edit: I took a second look. Since you've mostly been posting legit comments, I've unbanned your account in the hope that yet-one-more warning may finally have some effect. But please stop violating the rules here, especially with flamewar and personal attacks—you have a long history of doing this and it's seriously not cool. No more of this, please. https://news.ycombinator.com/item?id=36765186 https://news.ycombinator.com/item?id=36765186 (July 2023) https://news.ycombinator.com/item?id=36384568 https://news.ycombinator.com/item?id=36384568 (June 2023) https://news.ycombinator.com/item?id=36384562 https://news.ycombinator.com/item?id=36384562 (June 2023) https://news.ycombinator.com/item?id=35361931 https://news.ycombinator.com/item?id=35361931 (March 2023) https://news.ycombinator.com/item?id=35234576 https://news.ycombinator.com/item?id=35234576 (March 2023) https://news.ycombinator.com/item?id=34850116 https://news.ycombinator.com/item?id=34850116 (Feb 2023)
        giraffe_lady 3y ago
        I truly don't see it on this one man, sorry. It's ok to read the mind of a commenter and dismiss their motivations as "fashionable to virtue signal" but not to point out that doing that is shitty? I am disputing the words of a comment that it's author chose to put there, but I am not attacking the author's character or motivations. This is a genuinely held opinion, not "flamebait" or w/e. Please help me understand what rule I violated here, because I truly cannot see it.

14 points by andy99 3y ago

jxf 3y ago

Unrelated, but this is the first time I've seen the *.onl TLD.

Bias is a function of perspective. Every entity capable of generating output will be doing so based on "knowledge of" some set of inputs, even if that "knowledge" is part of its "hardware." It seems people are still wasting energy discussing "bias" in generative AI, so hopefully this ends the discussion as such. Obviously it is not possible to generate something without starting from copying something, this is the function of memories. In a physical system, this is t=n-1. Therefore it is not possible for a generative AI model to be free of bias. These discussions always seem to carry a moral connotation by using the word "bias" and relating it to hot-topic biases like racial and gender bias but the reality is that there's nothing necessarily immoral about a generative AI model's biases. Likewise, there's nothing necessarily moral about "fixing" these biases, because the reality is that "fixing bias" in a generative AI model is nothing more than realigning its bias to be more in line with the operator's own biases. This is an important distinction to realize because it helps to put the priority of the work in perspective. With a more accurate perspective of the priority we're better able to identify inefficient use of societal resources, such as taxpayer funding, postdoc researcher hours, and hours of general discourse.

giraffe_lady 3y ago
It seems like you're saying it will always have a bias, and one bias is as good as another. But while there may be nothing necessarily immoral about bias existing in general, but there certainly can be something immoral about applying a bias in particular. People aren't specifically focused on the ones you dismiss as "hot button" because of their popularity, they are hot button because those biases powerfully shape the world we live in. It is immoral to reproduce them, and it is morally necessary to choose not to.
- djyaz1200 3y ago
  [flagged]
  - giraffe_lady 3y ago
    [flagged]
    - dang 3y ago
      Since we've asked you many times to stop breaking the site guidelines and you've continued repeatedly to do it (e.g. here and https://news.ycombinator.com/item?id=37203245 https://news.ycombinator.com/item?id=37203245), I've banned this account. If you don't want to be banned, you're welcome to email hn@ycombinator.com and give us reason to believe that you'll follow the rules in the future. They're here: https://news.ycombinator.com/newsguidelines.html https://news.ycombinator.com/newsguidelines.html. Edit: I took a second look. Since you've mostly been posting legit comments, I've unbanned your account in the hope that yet-one-more warning may finally have some effect. But please stop violating the rules here, especially with flamewar and personal attacks—you have a long history of doing this and it's seriously not cool. No more of this, please. https://news.ycombinator.com/item?id=36765186 https://news.ycombinator.com/item?id=36765186 (July 2023) https://news.ycombinator.com/item?id=36384568 https://news.ycombinator.com/item?id=36384568 (June 2023) https://news.ycombinator.com/item?id=36384562 https://news.ycombinator.com/item?id=36384562 (June 2023) https://news.ycombinator.com/item?id=35361931 https://news.ycombinator.com/item?id=35361931 (March 2023) https://news.ycombinator.com/item?id=35234576 https://news.ycombinator.com/item?id=35234576 (March 2023) https://news.ycombinator.com/item?id=34850116 https://news.ycombinator.com/item?id=34850116 (Feb 2023)
      - giraffe_lady 3y ago
        I truly don't see it on this one man, sorry. It's ok to read the mind of a commenter and dismiss their motivations as "fashionable to virtue signal" but not to point out that doing that is shitty? I am disputing the words of a comment that it's author chose to put there, but I am not attacking the author's character or motivations. This is a genuinely held opinion, not "flamebait" or w/e. Please help me understand what rule I violated here, because I truly cannot see it.