I think you've misunderstood that example in the article. The AI isn't being asked to generate an image from the prompt, it's being asked to match the similar

I think you've misunderstood that example in the article. The AI isn't being asked to generate an image from the prompt, it's being asked to match the similar prompts to the different images. Winoground is basically a reading-comprehension test suite, which links back to the point made in the article that AI can't handle non-typical language precisely because it lacks reading comprehension (or any semantic model of language.) As the article points out, human runs of Winoground manage to match the vast majority of prompts to the correct image, so it's not a question of atypical language being too hard to understand. You may want to also read the author's other article[0] about the lack of semantic comprehension in AI models. 0: https://garymarcus.substack.com/p/horse-rides-astronaut https://garymarcus.substack.com/p/horse-rides-astronaut