Do AI-Generated Image/Art Algorithms Plagiarise? | Atomic Shrimp

Comments

• Even_Adder@lemmy.dbzer0.com ⁨1⁩ ⁨year⁩ ago

  Saying it learns exactly like humans is a straw man. This video’s not as neutral as he’d like you to believe.

  source

  • ProdigalFrog@slrpnk.net ⁨1⁩ ⁨year⁩ ago

    I don’t believe he’s advocating for that position, he’s attempting to illustrate the arguments either side on this debate make.

    source

    • Even_Adder@lemmy.dbzer0.com ⁨1⁩ ⁨year⁩ ago

      I believe he is. It’s a strategy called Fear Uncertainty and Doubt.

      source

      • -> View More Comments
• logicbomb@lemmy.world ⁨1⁩ ⁨year⁩ ago

  I admit I didn’t watch the video, since it is 17 minutes and I don’t have time right now, but I’ll just throw something out there that I think is a good rule of thumb.

  When you ask questions like this, “Do AI-Generated…” so and so, you can usually find a common-sense starting point answer by substituting, “Do human-Generated…” AI has the ability to plagiarize, and so do humans. AI has the ability to plagiarize even when it’s not asked to, and so do humans. Humans can even accidentally plagiarize, but it’s harder to say that AI does things accidentally.

  This rule of thumb doesn’t always work, but neural networks attempt to simulate the way human brains function. Obviously, there are some differences. But it’s close enough to get a starting point.

  It’s a complicated situation. A complicated question.

  source

  • ProdigalFrog@slrpnk.net ⁨1⁩ ⁨year⁩ ago

    The crux of this video is that the creator of it didn’t have a firm stance on either side of that argument until, during one bout of generations, Bing produced an almost exact replica of the "Is this _____?"meme. It crossed the threshold of ‘inspiration’ into out-and-out copying, and he provides a theory as to how that happened.

    Ultimately, he now feels uncomfortable using AI generations in any sort of commercial endeavor due to that risk of plagiarism.

    source

    • Even_Adder@lemmy.dbzer0.com ⁨1⁩ ⁨year⁩ ago

      Overfit images are few and far between in a well-trained, model. They are a product of error, and in reality you’d just be asked to eliminate the offending asset before anything major would happen.

      source