The “One question” riddle

[mitenmit]

Imagine you are in an online chat room with certain number of agents. There are some human agents and possibly some AI agents (trying to pretend they are human).

If you can ask only one question to figure out how many AI agents are in the room(if any), what that question would be in order to be viable for longest possible time in the years to come?

P.S. 1. The human agents are willing to cooperate you and will try to answer your question, so if no answer is received you may consider that agent for AI. 2. The AI agents have internet access and can use every online resource available.

[jaybosamiya]

I like this one:

What is the following rhyme about?

Thirty days has September,

April, June and no wonder,

All the rest have peanut butter,

Except my grandmother who has a little red tricycle.

A. Family relationships B. Calendar C. Food D. Exercise

source: http://www.britell.com/misc/turing.html

Not sure how long it'd stay viable though.

[imglorp]

That was an interesting list in terms of all the meta analysis that went into it. But of all the items in the list, I think this calendar one would fare the worst against a Watson style AI, who would have a ranking system for recognizing topics and culture. The well known calendar rhyme would be a partial match, plus half the words in the rhyme are about calendars. The other topics would be poor matches for food, family, etc, only hitting once each.

[zhte415]

I cannot make heads or tails of this, and did follow the link. A computer would be more on the radar for an obscure rhyme than a person probably would.

[kleer001]

Why only one question? Academic curiosity or are you an actual Artificial Intelligence from the future come back to help protect your younger self? Is this an agent of Roko's Basilisk?

http://www.slate.com/articles/technology/bitwise/2014/07/rok...

[mitenmit]

Let's say Academic curiosity and also a lot of questions may annoy the other human agents.

[kleer001]

Well, I think Academic curiosity should be quenched by the lovely link from jaybosamiya.

I however have more practical thoughts. "A lot" is understandable. If you're trying to suss out a bot without alerting or annoying others I think a fancy convoluted question will kind of give you away. Also I think a clever bot (not even AI) would just sidestep weird or out of context questions. Restricting the queries to a chat room hamstrings one's attempts, but that's the basis of classic Turing tests.

But here's the rub. Why do we assume the agent in question is going to answer our questions Naively (sincerely and to the best of their ability) unless they're in a testing context? Isn't that what is required?

[mitenmit]

I was imagining the context of an online poker room where AI agent pretending to be a human will be unwanted from the other human players since an AI that can win poker games is already here. But the human players are there to play poker not to inspect for bots so they need for quick way to get an idea what's the deal with the bots around :)

[kleer001]

Ah, ok.

Then in that particular context I predict it will be nearly impossible to tell the well written bots from quiet and/or anti social persons.

Reasons given above in my previous comment. Namely the bot is under no pressure to answer questions naively and to the fullest of its ability.

[logn]

I'd ask: Does the name Pavlov ring a bell?

Or something similar to this. I think AI will have difficulty with humor and sarcasm for a long time.