OpenAI has implemented a multi-tiered safety system for its AI model, DALL·E 3, to prevent the generation of harmful or inappropriate imagery. The system includes safety checks that run over user prompts and the resulting images. OpenAI has also worked with early users and experts to identify and address gaps in the safety systems, particularly in cases of graphic content and misleading images. Steps have been taken to limit the likelihood of generating content resembling the style of living artists or public figures, and to improve demographic representation in the generated images. OpenAI encourages user feedback to help improve the system and ensure responsible deployment of AI. Additionally, the company is researching a provenance classifier that can identify whether an image was generated by DALL·E 3 with over 99% accuracy in internal testing. However, further collaboration and development are required to make definitive conclusions and address the challenge of identifying AI-generated content.

DALL·E 3 is an advanced artificial intelligence system that specializes in generating images based on text prompts. It is an upgraded version of DALL·E 2, with improvements in caption fidelity and image quality. In order to ensure its safe and responsible deployment, the team behind DALL·E 3 conducted rigorous testing and evaluation. One key aspect of this process was external expert red teaming, where external experts were invited to assess the system and provide feedback on potential risks and flaws. The team also identified and evaluated key risks associated with the model and implemented mitigations to minimize these risks and prevent unwanted behaviors. The aim of this thorough preparation was to ensure that DALL·E 3 can be deployed securely and reliably, without any adverse consequences.

OpenAI is introducing new voice and image capabilities in ChatGPT, allowing users to have voice conversations and share images with the language model. These features provide a more intuitive way to interact with ChatGPT. For example, users can snap a picture of a landmark while traveling and have a live conversation about it, or take pictures of their fridge and pantry to decide what to cook for dinner and obtain step-by-step recipes. Additionally, users can help their children with math problems by sharing images of problem sets, enabling the model to provide hints and guidance. These voice and image capabilities will be rolled out to Plus and Enterprise users over the next two weeks. Voice functionality will be available on iOS and Android platforms through opt-in in the settings, while image features will be available on all platforms.