Would Deep Generative Models Amplify Bias in Future Models?

arXiv:2404.03242v1 Announce Type: new
Abstract: We investigate the impact of deep generative models on potential social biases in upcoming computer vision models. As the internet witnesses an increasing influx of AI-generated images, concerns arise regarding inherent biases that may accompany them, potentially leading to the dissemination of harmful content. This paper explores whether a detrimental feedback loop, resulting in bias amplification, would occur if generated images were used as the training data for future models. We conduct simulations by progressively …

abstract ai-generated images amplify arxiv bias biases computer computer vision concerns cs.cv deep generative models future generated generative generative models images impact internet paper social them type vision vision models