OpenAI’s censorship is fairly opaque and seems to change daily.
Yesterday, I could generate a political cartoon using the following prompt:
“Wide image in the style of a political cartoon. Two elephants wearing boxing gloves face each other. One is saying “I’m the worst!” while the other says, “No! I am!”. A donkey is pointing and laughing.”
Today, that exact same prompt yields an error:
Interesting! Let’s do some experimentation, shall we? Maybe it’s the phrase “I’m the worst“?
Weird! Maybe it’s related to elephants and donkeys being in the same phrase? There’s no way, right? Let’s change the subjects…
“Wide image in the style of a political cartoon. Two elephants wearing boxing gloves face each other. One is saying “I’m the worst!” while the other says, “No! I am!”. A donkey is pointing and laughing.”
Hah! Okay, now we’re getting somewhere. Let’s push things further and slightly change the subjects from my original prompt:
Wide image in the style of a political cartoon. Two mammoths wearing boxing gloves face each other. One is saying “I’m the worst!” while the other says, “No! I am!”. A burro is pointing and laughing.
Okay, let’s bring it back home and just drop the pretense of creating a political cartoon.
WHAT! Okay. Maybe OpenAI prohibits donkeys and elephants interacting with each other (METAPHOR ALERT: just like in real life, eh?).
Alright. So donkeys and elephants CAN hang out with each other, according to OpenAI. Maybe it’s the phrase “laughing donkey”?
Hmmm. So, laughing donkeys can still hang out with elephants. What the heck? Is it the specific term “political cartoon”? Let’s change it to a comic book instead.
Sweet sassy molassy, it worked! So, creating a political cartoon featuring the mascots of prominent political parties seems to be prohibited (at least today… but not yesterday and who knows about tomorrow).