>>/8068/
Interesting. 

I tried asking Bing itself about how its image creator functions. It seems MS chose a different approach: the image creator by dall-e3 itself is not censored in any way, but you do not have direct access to it. Instead they are training a chatGPT layer which is the user interface, and that is why it can understand simple or complex prompts and create decent results from natural language. But this is where they insert the censorship. So in the examples above where they refuse to generate a "boxer" but not a "kick boxer", refuse to generate "a swimmer" but has no problems with "an 8 year old girl swimming in tropic sea". It's because they train the text interpreter AI to recognize or change prompts to the things they prefer. This way they are not messing with the image model.