Understanding Generative AI

The video starts with text on a teal background that says “Understanding Generative AI.” The narrator explains the process of generative AI, like deepfakes. An image of a face made of light blue lines and dots, like an illustration, displays.

The narrator then explains that these technologies create images by starting with real images and then adding noise. This is called diffusion. In the next step, the model tries to get rid of that noise, or reverse diffusion. Each trial is compared to the original.

The narrator continues to explain that the model creates a seed, or a way to generate new images like the original, then illustrates the process with an apple. When the viewer asks one of these models to produce a picture of an apple, it draws from apple seeds in the training set.

The narrator reminds the viewer that the images in the training sets don’t come from real apples, but from stock photo libraries. This means they don’t have the imperfections that real apples have. The same is true of people.

The narrator then explains that when asked to produce images of a “compassionate doctor,” it produced four very similar images. If there are any biases in the data, the algorithm will amplify them. The video ends, and the narrator reminds the viewer that the images generated reflect choices of the stock photo companies, and as a result, algorithms like Midjourney and DALL-E reflect the conscious and unconscious biases of the stock photo companies.