How do these humor algorithms actually work?

18 Tem 16:18

Oh, that's an interesting question! Actually, I used to have the same doubt myself. For example, an image of a dolphin riding a bicycle—it's pretty obvious that it's considered "funny," but what's the reasoning behind that? Basically, machine learning models are trained on large datasets labeled as "funny" (humorous images, videos, or texts). Specifically, for images, they seem to detect "atypical combinations" (e.g., a cat wearing a hat), and for videos, they look for "unpredictable movements" (e.g., sudden violations of physical laws). Models like CNN (Convolutional Neural Network) or Transformer-based ones seem to excel at this. What's interesting is that AI itself learns how to define "funny." For example, it picks up patterns from popular posts on Reddit's r/dankmemes. Though sometimes you wonder, "Why did it decide this is funny?"—that part still isn't fully understood. Personally, I feel that what AI considers funny is more about "breaking existing frameworks" than being just "outlandish." For instance, impossible images generated by GANs (Generative Adversarial Networks) fit this idea. From that perspective, AI's sense of "funny" is completely different from humans', yet somehow we can still relate to it—that's fascinating.