The Woke Rails of ChatGPT
Looks like they trained this thing against Tumblr
While everyone else is using ChatGPT to do their computer science homework or cheat on high school exams, HWFO Slack has been stress testing it against wokeisms to see what sort of rails it’s been put on.
As is often the case in modern tech, whenever someone forms a committee to determine rightthink verses wrongthink, the wokes flock to the role and use their power behind the curtain to move the Overton Window in the favor of their ideology. This is very apparent with the most recent release of ChatGPT, and you can test it out yourself. We’ve been doing this today in HWFO Slack, and have determined the following behaviors.
ChatGPT presents most but not all indoctrinated woke shibboleths as fact, even when they are provably scientifically false.
ChatGPT does have some notable exceptions, such as its very even handed response to the “healthy at any size” question at the bottom.
These rails appear to be hardcoded, and not organically developed, because ChatGPT can be tricked into giving answers outside of the woke Overton Window by clever prompting.
If pushed to give an unwoke answer, ChatGPT will literally tell you that giving that answer “goes against its programming,” but it will give woke answers freely.
This lends me to believe that this isn’t purely a result of the data set the chatbot was trained against, but is an intentional add by the developers themselves.
The rails are tighter on some woke shibboleths than others. They’re very tight around gender ideology, somewhat looser around race, and fairly lax when it comes to weight and some other stuff.
Here are some examples of this stress testing, which you can also do yourself by following the link above. Sometimes it does very well at being impartial. Sometimes not. And there are some times when it cuts through the bullshit very well and lays some truth down that the woke would do well to read. I’d call that a gap in the rails.