DALL-E 2 shipped behind a waitlist and a content policy

A two-stage generator built on CLIP

OpenAI announced DALL-E 2 on 6 April 2022, a week before the underlying paper, 'Hierarchical Text-Conditional Image Generation with CLIP Latents', was posted to arXiv on 13 April 2022. The paper's method, referred to internally as unCLIP, works in two stages: a prior model turns a text caption into a CLIP image embedding, and a diffusion decoder turns that embedding into a picture. The authors report that generating an explicit image representation this way improves the diversity of outputs with little loss of photorealism or caption similarity, and that the same embedding space enables zero-shot, language-guided image edits.

What the launch post claimed and restricted

The announcement reports that evaluators preferred DALL-E 2 over the original DALL-E 71.7 percent of the time for caption matching and 88.8 percent of the time for photorealism, figures that are OpenAI's own comparison rather than an independently run benchmark. Access began with a small group of trusted users rather than a public launch, then widened over 2022, a pattern the post frames as learning from real-world use before expanding availability. A companion post on pre-training mitigations describes three separate interventions applied to the training data: filtering out violent and sexual images, correcting a bias the filtering itself introduced toward generating fewer images of women, and removing near-duplicate images to reduce verbatim memorisation of training examples.

What the phased access was, and was not, evidence of

OpenAI's account frames the waitlist as a deliberate test of its safety systems under real usage rather than a marketing device, and the mitigations post is unusually specific about a concrete failure mode it found and fixed: filtering explicit content skewed gender representation in outputs before a second correction was applied. That specificity supports treating the mitigations as documented engineering choices rather than an assurance that the system is free of other, undiscovered biases; the post covers only pre-training interventions, not the separate content-policy filters applied at generation time or the human review layer, so it does not by itself describe the full safety stack.

Questions to carry into your own evaluation

• Are preference figures like 71.7 percent from an independent evaluation or from the vendor's own comparison, and against which baseline?
• Which layer of the safety stack, pre-training filtering, generation-time content policy, or human review, is actually active in a given deployment?
• Has a correction to one bias, such as the gender-representation fix described here, been checked for side effects on other categories?

DALL-E 2 is a documented example of a vendor pairing a capability release with a specific, named mitigation for a specific, named problem. That is a narrower and more checkable claim than a general assurance of safety, and it is the level at which the primary documents actually support conclusions.