OpenAI disclosed modifications to DALL·E 2's image generation system aimed at reducing demographic bias and improving content safety [source]. The changes, implemented in July 2022, alter how the model interprets prompts that do not specify gender, race, or other demographic attributes.

The company reported that DALL·E 2 previously exhibited bias when generating images of people from underspecified prompts. Requests such as "CEO" or "lawyer" would disproportionately produce images of men, while prompts like "nurse" or "flight attendant" skewed toward women. The model also showed geographic and racial biases in its default outputs.

OpenAI's intervention involves automatically rewriting certain prompts before they reach the generation model. When users submit requests that could depict people but lack demographic detail, the system now appends descriptors intended to increase demographic diversity in the results. The company stated this technique applies to a subset of prompts and does not affect requests where users explicitly specify characteristics.

The disclosure also outlined safety mitigations. OpenAI expanded its content policy filters to block generation of public figures by name and implemented additional restrictions on violent, hateful, or adult content. The company acknowledged these filters may produce false positives, blocking legitimate requests.

This represents a documented case of a provider modifying model behaviour through prompt rewriting rather than retraining. Users submitting identical prompts before and after the change would observe different outputs, though OpenAI did not publish side-by-side comparisons. The intervention introduces a layer between user intent and model execution that may not be apparent to users unfamiliar with the system's architecture.