We use a multi-layered safety system to limit DALL·E 3’s ability to produce potentially harmful images containing violent, adult, or hateful content. Safety checks are performed on the user’s prompts and the resulting images before they are displayed to the user. We also worked with early users and a dedicated red team to identify and address gaps in safety system coverage that emerged due to new model features. For example, the feedback helped us identify edge cases of graphic content generation, such as sexual images, and stress test the model’s ability to produce convincing and misleading images.
As part of the work done to prepare for the rollout of DALL・E 3, we have limited the possibility of the model to generate content in the style of living artists or images of famous people, and we have focused on demographics across generated images. We also took steps to improve our presentation. For more information on the work done to prepare DALL·E 3 for widespread deployment, see DALL·E 3 System Card.
User feedback helps us continuously improve. ChatGPT users can share feedback with the research team by using the flag icon to notify them of unsafe output or output that does not accurately reflect the prompts they gave ChatGPT. Listening to our diverse and broad user community and understanding the real world is essential to developing and deploying AI responsibly and is core to our mission.
We are researching and evaluating an early version of the provenance classifier, a new internal tool that helps identify whether an image was generated by DALL·E 3. Initial internal evaluation showed greater than 99% accuracy in identifying whether an image was generated by DALL·E 3. If the image has not been modified, the image was generated by DALL·E. 95% or more, even when the image undergoes common types of modifications such as cropping, resizing, JPEG compression, or when text or cropping from the actual image is superimposed on a small portion of the generated image. Accuracy is maintained. Despite such strong results in internal tests, the classifier only indicates that the image was likely generated by DALL·E and cannot yet make a final conclusion. This provenance classifier could be part of a variety of technologies that help people understand whether audio or visual content is generated by AI. This is a challenge that requires collaboration across the AI value chain, including the platforms that deliver content to users. We hope to learn a lot about how this tool works and where it can be most useful, and refine our approach over time.