Russian Chatbot Passes Turing Test (Sort of)

CurtMonash writes "According to Ina Fried, a chatbot is making the rounds that successfully emulates an easily-laid woman. As such, it dupes lonely Russian males into divulging personal and financial details at a rate of one every three minutes. All jokes aside — and a lot of them come quickly to mind — that sure sounds like the Turing Test to me. Of course, there are caveats. Reports of scary internet security threats are commonly overblown. There are some pretty obvious ways the chatbot could be designed to lessen its AI challenge by seeking to direct the conversation. And finally, while we are told the bot has fooled a few victims, we don't know its overall success rate at fooling the involuntary Turing "judges.""

Comments

Re:WTF? This is not even a Turing test.

[cbart387]

It is not a turing test. A turing test is when the the judge is trying to figure out if the 'chatbot' is a human or AI program. This story is about people under the assumption that it is a human.

The key part of the turing test, to me, is that the judge must know they are engaged in the test. The best example of this is Eliza (read about it). To someone critically examining it, it does not past the turing test. To someone expecting a therapist, most of its responses do make sense. The point is that if you're not trying to trip up the chatbot it's not hard to fool someone.

Re:The ever-rising bar on true AI

[Haeleth]

Why is this post being modded up? It's a lovely example of the straw-man fallacy, but that hardly deserves Insightful moderations.

This chatbot has won the Turing test for a segment -- perhaps a gullible/dumb segment -- of the human population.

No it hasn't. It has convinced a gullible/dumb and unsuspecting segment of the human population that it is a human, which is not unimpressive in its own right, but that isn't the same as passing the Turing test, which requires that the examiner be conversing with a human and a computer at the same time, to be fully conscious of this fact, and to be deliberately trying to determine which is which.

Now we have chatbot that can fool some people some of the time, so the bar has been raised on "true AI" to say that computers can't fool expert suspicious Turing test judges. This too will fall.

Um, no. Nobody with a clue has ever claimed that a chatbot that is capable of convincing any human being whatsoever that it is a human represents true AI. The bar has always been set at fooling Turing-test judges, and the Turing test has been fixed in its current form for decades.

Indeed, it's easy to show that fooling some people some of the time doesn't require anything even approaching AI. Consider a bot that simply repeats a set of ten sentences in a fixed order: if those sentences were chosen well enough, then some people might easily believe that they were having a real conversation. But I really don't think you'd argue that a bot that simply repeated a set of ten sentences in a fixed order displays any sort of intelligence, no matter how many unsuspecting people happen, by random chance, to feed it lines that cause its responses to look relevant.

Turing probably was not serious about this test

[EmbeddedJanitor]

If you've studied Turings work much, you'd probably come to the conclusion that he never seriously proposed that the Turing Test would be a practical way to test machine intelligence.

Turing was a mathematician, which came through in all his thinking, including devising the Turing Test. When faced with questions like "can a machine ever be intellignt?" it is virtually impossible to answer this directly because, firstly, how do you define intelligence; and secondly,how do you measure intelligence?

Mathematicians **hate** imprecise questions because they cannot be proven or answered satisfactorily.

When faced with this problem, Turing used the well loved mathematical method of reductio ad absurdum: if you cannot tell the difference between a human and a machine, then it is absurd to claim the human is intelligent but the machine is not. That neatly sidesteps all the impossible to answer questions like the precise definition of intelligence. Typical mathematician wriggle out move.

Is the Turing Test practical? Well perhaps not. Machine intelligence (whatever that means) can be useful without the machine holding a conversation with you. Annoyingly it has soaked up a lot of effort with people building talkbots instead of getting on with more practical aspects of machine intelligence.