DALL-E 2 implements pre-training mitigations to reduce risks associated with powerful image generation models, including filtering out violent and sexual content.
The filtering process can amplify biases in training data, so techniques are used to mitigate this effect and prevent the model from being more biased than the unfiltered version.