Hi,
I’ve recently been experimenting with Sora and was surprised by some of the images it generated when prompted with satirical or historically charged scenarios related to WWII.
For example, I asked for mock children’s book covers in the Conni style (a German children’s book series) with titles like:
- Conni as a U-Boat captain
- Conni as a Stuka pilot
- Conni works at the WWII German secret service
After seeing that some outputs included swastikas, I began explicitly asking to avoid them—partly to feel more comfortable sharing the results—but the model sometimes accidentally included them anyway. I also noticed that Sora would sometimes use specific terminology like the name of the Nazi secret police, which I then asked it to generalize to “secret service.”
Even without the explicit symbols, I can imagine that many people might find some of these images tasteless or offensive—potentially coming across as “glorifying” rather than merely dark satire.
To be clear, I’m not arguing that this kind of content should or shouldn’t be allowed. Personally, I support minimal censorship and believe historical satire should be possible. But I’d also like OpenAI to be able to make those boundaries consciously, not by accident. I’m wondering whether these results slipped through unintentionally or are considered within the scope of current policy.
Two main questions:
- Is this kind of WWII-themed satire currently within policy bounds, or did I accidentally circumvent intended filters?
- Is there a preferred mechanism for “neutral” feedback—i.e., “I’m not offended, but you might want to review this”? Both the thumbs-down and report functions feel a bit mismatched for that use case.