ETHICS, PRIVACY & ORIGINALITY
As a new technology, we are still grappling to understand the extent and nature of bias and privacy risks associated with text-to-image models. While we know the capabilities of these models are impressive, it's important to recognize that they could be magnifying and mirroring societal biases on a global scale. With AI art models trained on private datasets or data crawled from the Internet, there's a possibility that they could produce images embodying or purporting harmful stereotypes.
As an example, the project thispersondoesnotexist.com, which generates a hyper-realistic portrait of a person who never existed, has been widely criticized for failing to generate people with black skin color, which alludes to issues with the training data used to create the model. (It’s worth noting that the project was spun up as an online stunt and demonstration to build awareness regarding the powerful capabilities of AI rather than as...