Asuka Ishii, Kazuki Nagata: Attraction of image
Artist(s):
Title:
- Attraction of image
Exhibition:
Category:
Artist Statement:
The fight against AI bias is progressing every day. Compared to measures against racial and gender inequalities, biases created by the Internet-dependent structure of datasets have not yet been thoroughly discussed.
The Internet is overflowing with convenient information that people expect. The images of tourist attractions in the dataset are obtained from social networking sites and tourist guides, and they always have blue skies and smiling people. The AI-generated images that learn from these datasets are not the reality itself but rather contain biases of the “place-ness” people want.
This work reveals that AI learns “place-ness” by comparing an AI-generated image of Enoshima island, a famous tourist attraction in Japan, with a photo taken on the actual Enoshima and edited for color and size to make it look exactly like that. Furthermore, by uploading these groups of images to Instagram and again creating a situation where they can become training data for the AI, we create the possibility of accelerating both our ideal image of Enoshima and the bias of AI. This work reveals the tip of the iceberg of the new bias created by the structure that images on the Internet are indispensable for the learning of large AI models.
