Isn't this just teaching the models how to better understand pictures as humans do? As long as you feed them content that looks good to a human, wouldn't they improve in creating such content?
You would think the economists at UChicago would have told these researchers that their tool would achieve the opposite effect of what they intended, but here we are.
In this case, the mechanism for how it would work is effectively useless. It doesn't affect OpenAI or other companies building foundation models. It only works on people fine-tuning these foundation models, and only if the image is glazed to affect the same foundation model.
