ChatGPT summoned up a conversation of improvements to this:
If I were summarizing your suggestion, it would be something like this:
Improve image generation block feedback for iterative creative workflows.
When an image request is blocked by automated safety systems, provide users with high-level, actionable reasons rather than a generic refusal. For example:
- Indicate the primary contributing concepts (e.g., “medical pathology setting”, “undead combat”, “graphic injury cues”).
- Distinguish between a likely policy violation and a low-confidence or automated classifier decision.
- Offer general guidance on which aspects of the prompt could be reworded without revealing internal detection mechanisms.
This would significantly reduce user frustration, especially for users iterating on artwork, character consistency, RPG campaigns, comics, or other long-running creative projects where a single blocked generation can derail hours of work.
I also think one of your observations is particularly important:
A continuity instruction should not appear to be the reason a prompt is rejected.
When a user changes only something like:
- “keep the same six characters,”
- “preserve their appearance,”
- “Dean uses the katana, Estobal uses ninjutsu,”
and the next request is blocked, the current feedback makes it seem as though the continuity instruction caused the rejection. Even if that’s not actually what happened, the user has no way to tell. Better feedback would make the process much less confusing.
Your point about distinguishing between:
- “This request violates policy,” and
- “This request was blocked by an automated classifier with limited confidence,”
is also a strong UX suggestion. Those are different situations, and users experience them differently.
Thanks for taking the time to explain the problem so clearly. It’s thoughtful feedback grounded in a real workflow, and it’s exactly the kind of insight that can help improve the experience for people using ChatGPT for iterative creative projects.