Eye to AI by Politecnico di Milano

The future is here: text-to-image AI is now a reality, enabling us to create images from simple text descriptions. Technologies like DALL-E and Stable Diffusion are rapidly spreading and becoming part of social culture. However, beneath the surface of this seemingly uncomplicated and entertaining technology lie significant ethical concerns that remain hidden from the public eye. Eye to Al brings to light the biases and stereotypes on climate change that txt-to-img AI inherits from society, using the machine itself as a tool to expose its inherent issues. The project aims to raise awareness on both climate change imagery and the myth of neutrality in artificial intelligence.
Using Stable Diffusion, we generated 520 images (through prompts ranging from climate change-related topics) and organized them into clusters based on observed patterns.

The outcomes can be explored through two main sections of the website: the Bias Catalogue and the Prompt Explorer.
The Bias Catalogue was designed to provide users with an overview of the content while highlighting distinctions among identified patterns. When exploring each bias, users can navigate through the images, zooming in and out for a more meaningful experience. These interactions offer multiple perspectives on the images: viewing them collectively from a distance serves to underscore the consistency of the biases, while observing them up close unveils critical details for comprehending these biases.
The Prompt Explorer showcases all the images collectively, emphasizing the relationship between the original prompt and the observed biases. This visualization depicts how these biases are distributed throughout the dataset. To convey this correlation effectively, we chose an exploratory tool that allows users to zoom in and out of images and freely navigate within the space. Furthermore, users have the option to filter images by bias, providing a rapid overview of bias concentrations within specific prompts.