5 ms·

Aren't there IRC bots already that successfully convince conversation partners to give them their credit card numbers?

lmm 11y ago
There are some, with certain (low) success rates. Expectations on IRC are weaker than those for an in-person conversation though; many people are inattentive or have limited English skills. I read an interesting excerpt from http://www.amazon.com/The-Most-Human-Artificial-Intelligence/dp/0307476707 http://www.amazon.com/The-Most-Human-Artificial-Intelligence... , arguing that the Turing test would become much more interesting once the humans started playing to win.
cousin_it 11y ago
Yeah. There are also bots that bring people to orgasm, and bots that hypnotize people over a text channel. I suspect that we might get bots that win AI-box experiments (convince a human gatekeeper to type the phrase "I let you out") long before we get "true" artificial intelligence. People are just too exploitable.
- schoen 11y ago
  Wow, having an AI-box chatterbot would be quite a thing. This may not be exactly the site for this discussion, but... When Eliezer did the famous AI-box experiments, there was a precondition that the human is supposed to accept claims by the AI to have demonstrated specific abilities (like curing cancer, solving math problems, or disclosing technical data about itself). Is it worth imagining a "nightmare difficulty" AI-box experiment where humans don't have to take the AI's word for it about anything outside of the text channel? Are there humans (or will there be chatterbots) that can win under those conditions too? How about against a more diverse set of conversation partners, with a wider range beliefs and attitudes toward AIs? How about if the AI's tactics don't have to be kept secret? (I'm not trying to minimize the significance of what Eliezer accomplished in those demonstrations, just curious about the prospects for even-more-difficult variants.)

lmm 11y ago

There are some, with certain (low) success rates. Expectations on IRC are weaker than those for an in-person conversation though; many people are inattentive or have limited English skills. I read an interesting excerpt from http://www.amazon.com/The-Most-Human-Artificial-Intelligence/dp/0307476707 http://www.amazon.com/The-Most-Human-Artificial-Intelligence... , arguing that the Turing test would become much more interesting once the humans started playing to win.

cousin_it 11y ago

Yeah. There are also bots that bring people to orgasm, and bots that hypnotize people over a text channel. I suspect that we might get bots that win AI-box experiments (convince a human gatekeeper to type the phrase "I let you out") long before we get "true" artificial intelligence. People are just too exploitable.

schoen 11y ago
Wow, having an AI-box chatterbot would be quite a thing. This may not be exactly the site for this discussion, but... When Eliezer did the famous AI-box experiments, there was a precondition that the human is supposed to accept claims by the AI to have demonstrated specific abilities (like curing cancer, solving math problems, or disclosing technical data about itself). Is it worth imagining a "nightmare difficulty" AI-box experiment where humans don't have to take the AI's word for it about anything outside of the text channel? Are there humans (or will there be chatterbots) that can win under those conditions too? How about against a more diverse set of conversation partners, with a wider range beliefs and attitudes toward AIs? How about if the AI's tactics don't have to be kept secret? (I'm not trying to minimize the significance of what Eliezer accomplished in those demonstrations, just curious about the prospects for even-more-difficult variants.)