How can an AI demonstrate purely through chat that it is an AI, and not a human?

post by hugh.mann · 2020-10-07T17:53:07.715Z · LW · GW · 1 comment

This is a question post.

Contents

  Answers
    5 Charlie Steiner
None
1 comment

In the latest William Gibson novel Agency, a bunch of human characters are constantly told exactly what to do by a superhuman AI in order to have everything turn out for the best. They accept the notion that this character, who they only talk to over their phones, is an AI, when it seems like it could just as easily be a fictional construct created by a person or organization. There's one part of the book where a character is convinced off-page that the AI is a superhuman AI by talking to it for a bit, although we're not privy to the conversation.

What I'm wondering is, how could a superhuman AI best prove that it was a superhuman AI through chat? Is there a more or less definitive way to do it? Is there a fast way to do it?

Answers

answer by Charlie Steiner · 2020-10-07T18:58:42.067Z · LW(p) · GW(p)

If we're sending text, then it can prove speed superintelligence by, e.g. simply answering many complicated questions all sent at once.

If we have to use voice, we can still try to ask hard questions and get fast answers, but because of the lower rate it's hard to push far past human limits. I can ask some pretty darn hard trivia questions with unambiguous answers, but someone with very good googling skills might be able to fool me.

One cute method might be to do something superhuman with the transmitted audio. For example, if it could tell how a die fell by listening to the sound of it rolling, read off handwritten digits from the faint sounds of writing, or typed letters from the acoustics of the keyboard.

comment by hugh.mann · 2020-10-07T23:31:23.026Z · LW(p) · GW(p)

Some of these things are falsifiable, for example people have already made programs that can detect letters from acoustics, and trivia questions would similarly be pretty easy, but the idea of asking a question that requires complicated processing to find the answer and requiring a superhumanly fast response seems like it might be the best. The trick would be that it would have to be a question that couldn't have been answered in advance (question can't be predictable, so that a person or organization could have prepared answers).

So it sounds like you'd need to be able to generate a question that 1. requires human-level or above reasoning, including generalization 2. is too complex for a human to quickly answer 3. cannot be predicted in form or content to an extent that would allow its answer to be cached in advance

comment by roland · 2020-10-09T12:50:43.143Z · LW(p) · GW(p)

If we have to use voice, we can still try to ask hard questions and get fast answers, but because of the lower rate itâs hard to push far past human limits.

You could go with IQ-test-type progressively harder number sequences.Use big numbers that are hard to calculate in your head.

E.g. start with a random 3 digit number, each following number is the previous squared minus 17. If he/she figures it out in 1 second he must be an ai.

1 comment

Comments sorted by top scores.

comment by hugh.mann · 2020-10-07T06:23:28.941Z · LW(p) · GW(p)

I remembered after writing this that there are also scenes where the AI provides people with schematics for advanced technology in return for labor; in these situations, it isn't trying to prove that it's an advanced AI though, and even if it were trying to use this as a proof there's no way to prove that the technologies it provides weren't actually developed by humans, even if they were advanced.