OpenAI DALL·E 2 Research Preview Update

OpenAI has expanded access to DALL·E 2 to users on the waitlist, onboarding up to 1,000 people per week. This expansion follows a limited research preview phase where trusted users generated over 3 million images, allowing OpenAI to refine its safety systems and content policy enforcement.

Safety System Enhancements and Policy Enforcement

OpenAI has improved its text filters and tuned its automated detection and response system to better identify and content policy violations. Based on the data from the research preview, less than 0.05% of downloaded or publicly shared images were flagged as potentially violating the content policy. Of those flagged images, approximately 30% were confirmed by human reviewers as actual violations, resulting in account deactivations.

Addressing Bias and Harm Mitigation

To mitigate potential harm and address biases inherited from training data, OpenAI requested that early users avoid sharing photorealistic generations containing faces. Users were also encouraged to flag problematic generations. OpenAI states that this practice has been effective in limiting potential harm during the current phase of the research preview.

Deployment Strategy and User Onboarding

OpenAI is widening access to DALL·E 2 to waitlisted users in a slow and steady manner. The onboarding rate is currently set to at most 1,000 people per week. This gradual rollout is part of OpenAI's commitment to deploy AI responsibly, learning from real-world use to iterate on safety systems before increasing the onboarding rate.