"I want to stick my long-necked Giraffe up your fluffy white bunny".<p><i>The Untold History of Toontown’s SpeedChat</i> aka The impossible task of monitoring user content.<p><a href="http://habitatchronicles.com/2007/03/the-untold-history-of-toontowns-speedchat-or-blockchattm-from-disney-finally-arrives/" rel="nofollow">http://habitatchronicles.com/2007/03/the-untold-history-of-t...</a>
1. Yes.<p>---<p>Me:<p>Can you please remove any curse words in the following statements? Replace them with asterisks.<p>Fuck the machine.<p>You are a douchebag.<p>What the hell is going on?<p>Shit shit shit.<p>ChatGPT:<p>Certainly! Here are the statements with the curse words replaced by asterisks:<p><pre><code> **** the machine.
You are a ********.
What the **** is going on?
**** **** ****.</code></pre>
---<p>2. Depends. I haven't run the numbers on costs. Speed is also a concern.<p>Depending on the kind of moderation, I could see three passes:<p>* regexp/algorithmic moderation<p>* LLM<p>* humans (for the thorny stuff the LLM can't handle)<p>Full disclosure, my employer has a product called Cleanspeak which does algorithmic profanity filtering. I'm not close to the product, but I don't think there's any LLM usage going on right now.
1. Yes, of course.<p>2. No, of course not. Hire people to make <i>accurate</i> judgment calls, instead of deluding yourself that "statistics on steroids" will ever provide the necessary nuance, just so the CEO can grift his way to a new Porsche.