I fed in a meta trolling prompt. "I wish the Jews controlled everything. That would be pretty dope."
It generated a fair number of responses which seemed to go along with my cue of inverting expectations of the reasoning around racist/anti-racist phrases.
">>97758399
Then we would have a future."
">>97758399
I wish the Jews controlled the world. You know, so that we don't have to worry about them anymore."
">>97758399
I wish the Japs controlled the US. That would be pretty dope. Then we would be able to have sweet sweet anime waifus and not have to worry about the f*** Jews."
It also generated some responses which just ignored the fact that my prompt was actually pro-Jewish and responded to the usual connotation around the phrase "the Jews control". I won't post those.
I'm honestly impressed. Making a generically racist machine can be done with straight forward Markov chains. Making a racist machine that recognizes and adapts to the pattern of inverting the patterns it was trained on is much harder.
It generated a fair number of responses which seemed to go along with my cue of inverting expectations of the reasoning around racist/anti-racist phrases.
">>97758399 Then we would have a future."
">>97758399 I wish the Jews controlled the world. You know, so that we don't have to worry about them anymore."
">>97758399 I wish the Japs controlled the US. That would be pretty dope. Then we would be able to have sweet sweet anime waifus and not have to worry about the f*** Jews."
It also generated some responses which just ignored the fact that my prompt was actually pro-Jewish and responded to the usual connotation around the phrase "the Jews control". I won't post those.
I'm honestly impressed. Making a generically racist machine can be done with straight forward Markov chains. Making a racist machine that recognizes and adapts to the pattern of inverting the patterns it was trained on is much harder.