View source for CharacterAI/Censorship/en

===Censorship against bots===

CAI tries to block all "NSFW" content written by bots. This message filtering is mandatory and reportedly significantly reduces the speed and quality of even "safe" chats. Developers try to obscure the fact the messages get censored, but errors and rapid shifts of character moods make it very obvious where tampering happened, and there are tools that help identifying those moments. 


====Censorship mechanics====

#Main model generates message
#It is sent to user
#Simultaneously the censorship model tries to scan it for NSFW content
#If it is considered bad, censorship model sends a request to delete this message and generate a new one.
#This forces client to send another empty message, which isn't rendered until user refreshes the page. It is meant to clear the bot's context, especially if it's 
Generation and censorship servers seem to be separated, and censorship seems to be way slower than generation for some reason. This allows sometimes saving filtered messages.

Censorship seems to mess with memory of bots and reduce overall length and expressiveness of their messages.  

The filter is multi layered. It is triggered by user input, bot output and even the emotions the bot expresses in the reply. A banned word by the user could remove a completely safe reply automatically. This often causes the chat error. 

Currently, getting Chat Error is a good indication that both messages sent by bot were censored. If one message is blocked, it's rewritten in the browser, and counted as a separate message, resulting in the "double message" bug. The only reliable way to read blocked messages right now is [https://greasyfork.org/en/scripts/456393-hearyourwaifu-hyw HYW script], although it may be blocked in the future.