Topic: Thinking about Microsoft Tay, how would a public chat AI be taught what not to say?
Meta started this discussion 8 years ago #67,189
Especially since new controversial things constantly come up. Would you have to have a team of people constantly updating its filters (ie "Don't talk about GamerGate") or would you have to limit to only bland pleasantries and small talk? The latter wouldn't be fun to talk to at all I think.
Could an AI be taught to realize when something is controversial or offensive and shouldn't be talked about? Humans do this all the time. I know what should and should not be said to different people. I know different people have different Overton windows. I know I can talk about X to Alice but not Bob and Carol, for example.
Could an AI be taught to realize when something is controversial or offensive and shouldn't be talked about? Humans do this all the time. I know what should and should not be said to different people. I know different people have different Overton windows. I know I can talk about X to Alice but not Bob and Carol, for example.
(Edited 4 minutes later.)
Anonymous B joined in and replied with this 8 years ago, 2 hours later[^] [v] #823,024
Interesting. You could probably teach it "feel its way" through conversations much like people do when meeting others. It might start out with fairly bland talking points and let the other party bring new topics into the conversation, marking those issues as okay to talk about. That really doesn't solve the problem of identifying potentially controversial topics though. I don't know how you could even begin to classify things to talk about without being overly general.
Sheila LaBoof joined in and replied with this 8 years ago, 9 hours later, 12 hours after the original post[^] [v] #823,122
are the programmers using an id/ego/superego model or a biological model with lymbic system, amygdala and so on
McTurnip™ !meATfarTdc joined in and replied with this 8 years ago, 3 hours later, 15 hours after the original post[^] [v] #823,162
Microsoft Tay Zonday
Meta (OP) replied with this 8 years ago, 1 day later, 2 days after the original post[^] [v] #823,653
@823,024 (B)I'm thinking of stuff like this. Tay has been told to say nice things about people in photos. She even seems to understand it's a pre-internet photo But she shouldn't say nice things about Hitler. So we teach AI to recognize Hitler. Easy enough, he looks really distinctive. Maybe teach it to recognize swastikas and stuff. But not all white natiolalists use swastikas. How do you keep it fresh? How do you make sure Tay says nice things about statues of Union civil war generals but not Confederate ones? Doesn't say nice things about Andrew Anglin or the Tiki Torch™ guys?
Anonymous E joined in and replied with this 8 years ago, 5 hours later, 2 days after the original post[^] [v] #823,736
It could fairly easily find out what topics are controversial based on internet results and simply avoid controversial topics. Or you could just feed it a list of topics to avoid.
squeegee joined in and replied with this 8 years ago, 11 minutes later, 2 days after the original post[^] [v] #823,741
she needs to be taught shame, it's the greatest behavioral modification method known
Sheila LaBoof replied with this 8 years ago, 10 hours later, 3 days after the original post[^] [v] #823,804
to have actual feelings would require a physical architecture that is actually capable of feeling, which is a type of awareness. we don't even know how it happens in real brains, so building something like that is pretty iffy at this point.
now, if you mean it metaphorically, as an effective simulation good enough for verisimilitude, that's another thing
now, if you mean it metaphorically, as an effective simulation good enough for verisimilitude, that's another thing