We use a multi-tiered security system to restrict DALL·E 3’s capability to generate doubtlessly dangerous imagery, together with violent, grownup or hateful content material. Security checks run over consumer prompts and the ensuing imagery earlier than it’s surfaced to customers. We additionally labored with early customers and skilled red-teamers to establish and tackle gaps in protection for our security techniques which emerged with new mannequin capabilities. For instance, the suggestions helped us establish edge instances for graphic content material technology, comparable to sexual imagery, and stress check the mannequin’s capability to generate convincingly deceptive pictures.
As a part of the work accomplished to arrange DALL·E 3 for deployment, we’ve additionally taken steps to restrict the mannequin’s chance of producing content material within the type of residing artists, pictures of public figures, and to enhance demographic illustration throughout generated pictures. To learn extra concerning the work accomplished to arrange DALL·E 3 for broad deployment, see the DALL·E 3 system card.
Consumer suggestions will assist ensure we proceed to enhance. ChatGPT customers can share suggestions with our analysis workforce through the use of the flag icon to tell us of unsafe outputs or outputs that don’t precisely replicate the immediate you gave to ChatGPT. Listening to a various and broad group of customers and having real-world understanding is vital to growing and deploying AI responsibly and is core to our mission.
We’re researching and evaluating an preliminary model of a provenance classifier—a brand new inner software that may assist us establish whether or not or not a picture was generated by DALL·E 3. In early inner evaluations, it’s over 99% correct at figuring out whether or not a picture was generated by DALL·E when the picture has not been modified. It stays over 95% correct when the picture has been topic to frequent forms of modifications, comparable to cropping, resizing, JPEG compression, or when textual content or cutouts from actual pictures are superimposed onto small parts of the generated picture. Regardless of these robust outcomes on inner testing, the classifier can solely inform us that a picture was doubtless generated by DALL·E, and doesn’t but allow us to make definitive conclusions. This provenance classifier could turn into a part of a spread of methods to assist folks perceive if audio or visible content material is AI-generated. It’s a problem that can require collaboration throughout the AI worth chain, together with with the platforms that distribute content material to customers. We count on to study an ideal deal about how this software works and the place it may be most helpful, and to enhance our method over time.