An estimated 99% of farm animals in the United States are raised in concentrated animal feeding operations, or CAFOs. Often called factory farms, these facilities raise livestock in densely packed, enclosed spaces, with some housing tens of thousands of animals at one time.
[…]
In a new Technology in Society paper, Sheng and co-authors analyzed 4,800 independently generated images and found that ChatGPT defaulted to pastoral imagery when prompted to depict dairy and pig farms, even when they requested more realistic images. However, when researchers disabled the underlying prompt revision, the generated images depicted more intensive livestock farming.
well, to be fair, we trained most models away from non-sfw images, so ugly images get cleaned up to look nice. That’s a consequence of making image models that don’t do things you don’t want your kids to see
I was thinking the same. They’re probably weighted against gory, distressing images. Think of the headlines: “<LLM service> creates graphic slaughter house for children’s storybook project”
OLD MACDONALD NO!!!
It’s a little bit more specific than that, prompt rewriting is a step they added intentionally to remap prompts. So not exactly trained to make them less about reality, but actually changes your prompt to the image generator to be less about reality
Sheng found an explanation deep within the technical documentation for DALL-E 3, OpenAI’s text-to-image AI model for ChatGPT: An internal language model rewrites or expands user prompts to add more details before images are generated.
Yeah. If I asked you to draw a picture of a pig farm, you’d probably draw a barn and pigsty with pigs and a trough, too. Which way should it be weighted, if any?
I don’t know how someone can think “AI CAN obscure reality …”
Obviously the fundamental truth is “AI does obscure reality …” (Often this will be inconsequential, but that doesn’t make it less true).
(Edit: especially an academic, FFS)