How does the Turing test work, and why is it still debatable whether it has been passed or not?


How does the Turing test work, and why is it still debatable whether it has been passed or not?

In: Technology

Simple. You are put in a room with a 2 keyboards and 2 screens. On one end there is a person, on the other, a programm.
You ask both whatever questions you want, for how long you want.
The person is not trying to trick you. If you can’t figure out reliably which is the computer and which is the person, the programm has passed the turing test

There is no one “Turing test”. To understand why people talk about it, quick history lesson.

At the time computers were being envisioned by early pioneers, we had this idea that machines should be able to do computations. But what is computation? Multiple people, among them Turing, had their own idea of how to rigorously define what is computation, to have this theoretical background that makes sure computers can do any computation. Turing machine was Turing’s answer to this, and they formulated cool Turing-Churchill conjecture saying that yeah, this Turing machine idea probably encapsulates what we mean by computation.

Anyway, in this world where people are trying to define what computation even means, people wanted to know if these computation machines could think. And lots of philosophical debates started. What does thinking even mean, how do you detect presence of thought, is human brain uniquely capable of thought…

And Turing came up with this simple test to direct people towards something practical instead: if you cannot tell the difference between words produced by a computer that may or may not think, and words produced by a human, does it matter if computer thinks or not? From your point of view, it shouldn’t matter. So Turing simply proposed, lets test that then to see if computer thinks, if you can tell the difference.

For the most part, Turing test obviously has not been passed. You cannot have free-form discussion with a computer without illusion of thought shattering quickly. But if you restrict the test or scope of questions or length of interaction, you can get fairly good results for computer. And the thing is, there is no One True Turing Test. Turing test is an idea about testing humans and their ability to differentiate computers from humans rather than trying to assess computer thought directly. It doesn’t have one set of rules, you can do tests in that spirit in multitude of ways, and people have done so. All that matters is that ultimately a human is tasked to tell the difference between a human(that presumably is capable of thought) and a computer(the supposed ‘unknown’) from their output alone.

The Turing test is not a real test so it doesn’t really “work” in any way. It’s based on a weird game that bored British people used to play, called “the imitation game” (hence the name of the movie), where people would pretend to be of the opposite sex.

In this game, participants can only communicate using written notes and Turing devised a thought experiment analogous to this, but with humans and computers.

Since this was quite a while ago, and people back then had vastly different understandings of computers, the original test as posited by Turing is somewhat different than the modern versions you may have seen. Turing was absolutely an incredible genius and a giant of computer science, but his original idea for the Turing test has a certain naivete and is a bit weird, partly due to what computers were then and partly due to being inspired by that strange game.

The point is, the Turing test is more of a Turing thought experiment and doesn’t really have a rigorous definition or any hard numbers attached to it. It’s not meant to. It’s not intended to actually determine whether an AI is an AI or a person, but it’s meant to ask the question “can machines think?” in a metaphorical way since it was asked at a time where computers were far less understood and developed.

Now, lay people, journalists and storytellers absolutely love the notion of a test meant to tell AI and humans apart and just assume that’s exactly what the Turing test is.

There have been, and likely always will be, arguments over whether AI systems “actually think”. The Turing test is a sneak attack on the vulnerable flanks of that question. If an AI system can perform some activity (classically, engaging in conversation) sufficiently well that it is indistinguishable from a human performing the activity, it works, and it doesn’t matter if what it does is or is not “actual thinking”.

I believe it was Turing taking the piss, but it may have been simply a way clearly to express one of AI’s goals.

As laid out in the original paper, it’s not at all debatable: it hasn’t been passed. If for no other reason than you’d first have to play “The Imitation Game” a whole bunch of times in order to collect statistics on how often men can successfully imitate women.

The idea is simply that, absent telepathy, you can’t tell the difference between a computer that thinks, and a computer that behaves as if it thinks. All you ever really have from a super complex system is what you can observe from the outside, so let’s observe the computer in conversation, lying.

It probably says something that led Turing figured “lying” was the most human of activities and the best one to test for when trying to see if a computer was thinking like a human.