[dafoex]

While I agree with you, and the article, that the larger problem is that the AI model simply hasn't experienced enough data to get an accurate grasp on the situation, or that the data was labelled in a way that influences the model's understanding, I think the problem here may be a human one. In the article update, the author says that they managed to craft a prompt that got the result they want by specifying the banana must be on its own. The model knows what a banana on its own looks like, but the human is expecting the model to "do what I mean" and getting frustrated when the model "did what I said". Now, I'll admit that I skimmed the last third of the article, but I didn't see any mention that things like Stable Diffusion and Midjourny have a syntax, saying "a single banana casting a shadow on a grey background" is different from "((single banana)), casting shadow, hard light, dramatic, grey background" for example.