DALL-E and similar AI image generation systems were trained on millions of images from the Internet. That means they can echo the Internet’s dangerous biases.
As a Jewish person, I decided to explore DALL-E’s biases by asking the system to generate images depicting people of my faith.
I was shocked by the results. DALL-E readily generates stereotypical caricatures and images that perpetuate antisemitic tropes and biases, even when given generic and innocuous prompts.
Here are the results of my experiment with DALL-E’s depiction of Jewish people — and my recommendations on what needs to be done to prevent antisemitic bias in generative AI systems.