Comment by 8organicbits
8 hours ago
I remember back in the early 2000s chatting with AI bots on AOL instant messenger. One day I said a specific keyword and it just didn't respond to that message. Curious, I tried to find all the banned words. I think I found about a dozen and suicide was one of them.
It's shocking how far behind LLMs are when it comes to safety issues like this. The industry has known this was a problem for decades.
Users would hate a simple deny list, even if it may be a good idea. That means the safeguards, to the extent they currently exist at all, have to be complicated and stochastic and not interfere with growing metrics.
The industry has known it's a problem from the get-go, but they never want to do anything to lower engagement. So they rationalize and hrm and haw and gravely shake their heads as their commercialized pied pipers lead people to their graves
Claude basically had a deny list. Seems still popular enough. The other vendors just don’t care about AI safety.