The article also mentions the development of a provenance classifier, an internal tool to identify whether an image was generated by DALL·E 3. The classifier has shown strong results in internal testing, with over 99% accuracy when the image has not been modified and over 95% accuracy when the image has been modified. However, the classifier can only suggest that an image was likely generated by DALL·E, and does not yet enable definitive conclusions. The tool is expected to be part of a range of techniques to help people understand if audio or visual content is AI-generated.
Key takeaways:
- DALL·E 3 uses a multi-tiered safety system to limit the generation of harmful imagery and has undergone extensive testing to identify and address gaps in safety coverage.
- Steps have been taken to limit DALL·E 3's ability to generate content in the style of living artists, images of public figures, and to improve demographic representation in generated images.
- User feedback is crucial for continuous improvement and users can share feedback with the research team to report unsafe outputs or outputs that don’t accurately reflect the given prompt.
- A provenance classifier is being developed to identify whether an image was generated by DALL·E 3, with early evaluations showing over 99% accuracy for unmodified images and over 95% accuracy for commonly modified images.