OpenAI publishes a system card for DALL·E 2

Alongside the limited early-access preview of DALL·E 2, OpenAI published a “system card” — a document assessing the risks and limitations of the image-generation model rather than its architecture or training method. It was released to the roughly 400 trusted testers who had access to the preview at the time.

The card grouped risks into several categories: explicit content (the model would not generate images from prompts naming violence or nudity directly, but could produce comparable imagery through visual synonyms — the card’s own example was requesting “ketchup” to depict blood on a person); bias in who and what the model depicted, including over-representation of white-passing subjects and Western imagery and reinforcement of gender and occupational stereotypes; harassment, since the model’s inpainting feature could edit real people’s photos in ways usable for bullying or blackmail; disinformation risk from photorealistic fabricated images; and the potential for economic disruption in photography, illustration and stock-image work. Listed mitigations included restricting the preview to vetted users, filtering both prompts and uploaded images, rate-limiting API access, requiring disclosure that outputs were AI-generated, and applying a visible watermark to generated images.

The document did not claim these risks were solved — it explicitly flagged that further red-teaming was needed on explicit content in particular. Its significance was less in any individual finding than in the practice it established: publishing a structured harms assessment as a companion document to a model release, ahead of wider availability, became the template OpenAI and other labs repeated for subsequent releases, most visibly in the GPT-4 system card the following year. OpenAI widened access to DALL·E 2 substantially three months later, adding further safety mitigations as the user base grew from hundreds to roughly a million.