When do you believe software AIs will truly be able to pass the Turing Test?

Question

When, if at all, do you believe that we will see AI systems that are capable of passing the Turing Test?

If a machine acts as intelligently as a human being, then it is as intelligent as a human being

This concept forms the basis of the Turing Test, a means to evaluate if an artificial intelligence is indistinguishable from a human being by observing its behavior and interaction with human beings. As yet, no system has been able to pass as human (although numerous human been mistaken for machines).

If you believe that we will eventually build such systems, I would be curious to know what approach you think will be successful.

If you don't think such a system is likely to be created any time soon, what do you think that obstacles are?

Answer 1

Six to eight weeks.

Answer 2

December 21, 2012

Answer 3

Jon Skeet is the only computer that passes the Turing Test.

Answer 4

Questions like this remind me of a quote a friend of mine had in college

4 years spent in AI research will really make you question your atheism

Answer 5

I once worked with - well, alongside - a guy who I swear was trying to pass the TT in reverse. You do get a lot of socially challenged people in this profession...

Answer 6

Ten years from now. And in ten years, I reserve the right to answer "10 years from now" again.

Answer 7

I don't think that the Turing test is necessarily a good test of AI as it will be broken based on a great graphics parsing algorithm...not a thinking computer. I think these concepts are totally different.

Answer 8

Several chat programs can fool actual people "Eliza" is probably the first and seems primitive now, although she can actually fool some people. And these programs continue to get more sophisticated .

Answer 9

I just want to note that the validity of the Turing test as a true test of intelligence has been under fire for quite some time (I believe since it was first stated!), so it's not necessarily recognized as the perfect and complete human-equivalent-intelligence test (though certainly still the most famous). Also, though people seem to think computers are all-powerful given enough resources and time, many problems are provably incomputable by a Turing machine - and problems far more practical than simply the well-known Halting Problem.

Isaac Asimov wrote what is perhaps the most famous sci-fi short story ever written called The Last Question. It dealt with a computer that exponentially increased in intelligence until being God-like. I found this discussion very interesting on why Asimov's computer could be considered an exaggeration due to incomputable problems.

Answer 10

If a machine acts as intelligently as a human being, then it is as intelligent as a human being

You're going to have to write a lot more conditionals to get that program working.

Answer 11

When we design learning computers, it turns out that they don't and we do.

Answer 12

The level of progress made on this front has slowed, but I think we will eventually reach the point where this is possible. Instead of directly creating the AI ourselves I think the human race could reach this goal indirectly through the form of the singularity.

The theory goes that through technical advancement we will eventually create a machine that is capable of re-designing an improved version of itself, this will result in a "intelligence explosion" whereby through quick iterations (of accelerating pace) this machine will very quickly surpass human intelligence.

A recent interview in new scientist with Ray Kurzweil, he stated that he believed that the singularity would occur sometime midway through this century.

Answer 13

When the judge is Turing....

Answer 14

At the current rate, > 100 years. It's important to make big predictions so that somebody can prove you massively wrong in the next year or so ;-)

Answer 15

We don't know enough about human intelligence to effectively simulate it. (It's at least arguable that we can't know enough about human intelligence, that intelligent beings can only understand things significantly less complex than they are.)

This means that we will get intelligent machines through some thoroughly empirical process using some sort of evolutionary algorithm.

This means that they will think considerably differently from humans, and will be distinguishable in a Turing test. However, neither will a human be able to pretend to be one of those AIs.

Therefore, if the AIs ever develop the ability to be indistinguishable from humans, they'll be a whole lot smarter than we are.

Answer 16

I suggest reading some Ray Kurzweil books. The Age of Spiritual Machines is great (from back in 1999) as well as The Singularity is Near (2005 I think). In both books, he predicts that in around 20 years we'll have computers than can perform as many calculations per second as many scientists feel that human brains can do. And then he believes that it won't be long after that before we start seeing true AIs.

I've also read that the size, complexity and shear computing power of all the computers connected to the internet could bring about some form of global computing awareness. I'm not sure if that would pass the Turing Test, but it'd sure be interesting!

My own prediction is we'll have human level AI within 30 years, as I expect the Singularity to happen within that time.

Answer 17

While this may seem a laudable goal, I question whether it's actually of any value. Do we really need another human?

The history of tool development has mostly been about extending the ability of humans to do thinks an unaided human can't do well: to lift heavy objects, move water uphill, travel faster than the fastest horse, breathe underwater, compute where a bomb will land if thrown just so, or unscramble a cunningly concealed message.

How about making an AI that can tell really good jokes, every time? Or one that can tactfully choose my wardrobe every morning? How about an AI that can govern a city without corruption or prejudice?

Aim higher please.

Answer 18

I don't believe it will ever happen.

(Sorry to the enthusiasts)

The full reason could be it's own book. The short answer is that a person can only understand intelligence that is either equal to or less than themselves. Anybody who is in "gifted" programs or especially genius level can tell you that.

You can say "That person is smarter than me" and sort of know that it is true, but to actually understand how or why means you would have to have equal intelligence to the thing you are understanding.

That said, to create an intelligent machine you have to actually comprehend more than the machine you create can comprehend. Just like any transfer of energy, the transfer of intellect will incur loss.

If it helps, take a step back for a little bit. How well do you understand your own intellect? If you are honest you have to say you do not understand it as well as you can use it.

Answer 19

Yes, absolutely, but not for the commonly understood reasons. The problem is that although the intended way for program to pass the Turing Test is to simulate or emulate human intelligence through the unique(for now) human activity of communication via human language, the Turing Test is flawed and allows a significant shortcut:

Instead of focusing on emulating intelligence, a program can instead focus on fooling humans, which it turns out is not all that hard (as an endless number of "Nigerian Finance Ministers" have proven). ELIZA is the example most often given to demonstrate this, but IMHO, PARRY is the much more significant case as the ELIZA concept is limited because of the difficulty of extending it to larger communication & knowledge spaces, however there isn't really anything about PARRY that would prevent it from being scaled up to a huge level.

That this is possible was demonstrated some years ago by a Turing Test, where the part of the "Computer" being tested was played by a program that was also intercepting the communications to and from the Humans under test and then simply copying an answer to a question that most closely matched the one that it was just asked. This (again IMHO) is a shallower, though much broader example of the PARRY approach.

The inescapable conclusion is that we could readily develop a program, today, that could handily pass most of the even moderately limited Turing Tests if we just put the time and money into it (note: most Turing Tests conducted today are not moderately limited, but severely limited tests). How could this most easily be done? As follows:

Create a version of the Google spiders that constantly scans the Internet for textual exchanges and then store and categorize the responses by semantic analysis of the questions (or initiating text) and store it in a specialized version of the Google databases. Next constantly invite people to participate in on-line Turing Tests as both testers and testees. Record and catalog the dialogs as before but giving more weight to these exchanges.

Next, participate in Turing Tests, scoring itself on how often testers guessed that it was a computer. Finally use the type of massive tree-branching weighting & look-ahead search and move-selection algorithms used by Chess & Checkers programs currently (adapted for semantic analysis and exchange or course).

Now our program is ready for the real public Turing Tests. And of course while it is doing that continue to run low-profile Turing Tests of its own around the world, feeding it's questions/prompts that it receives from the real testers to those other folks and storing their responses as a backup for more complicated and involved situations. And of course, apply the first principle of Chess, Checkers and PARRY: Take control of the engagement, thus limiting the other side's options.

So, in short: the Internet can pass the Turing Test.

Answer 20

Mr. Data could probably pass a Turing test, although not every time, and he's from the 24th century, I think. I think that qualifies as an authoritative answer.

Answer 21

I don't think computer programs have to get smarter to pass the Turing test. I think we just need a little bit more of a decline in human intelligence.

Answer 22

As long as AI researchers maintain the idea that human intelligence is based entirely within the brain, we'll make no progress. Psychologists have progressed beyond that idea, but AI researchers, as far as I've seen, seem stuck in that mindset, and are pursuing primarily brain based theories.

I've seen rather convincing arguments for the idea that human intelligence works a bit more like this: While a great deal of sophisticated algorithms are located within the brain, quite a lot of what we recognize as human behavior is the result of the brain, the body, and the environment all working in concert. If you wanted to create a computer that could convincingly simulate a human, the computer would need a body, an environment, and a childhood.

With enough advances in hardware, it's conceivable that a clever programmer could simulate the body,environment,childhood in a computer on a vastly accellerated timescale. A conversation could be staged by having a simulated AI researcher ask that simulated human to be an examiner in a turing test.

When the hardware gets good enough to do that, and the AI researchers get clever enough to do that, then we'll be 10 years off.

To contrast, imagine you just heralded in your first born child, a lovely 8 pound boy, and immediately extracted its brain, and put it in a box wherin the only stimulous the brain recieved was ascii encoded text. Just how intelligible would you expect that person to be immediately? How intelligible would it be after 10 years of that treatment? 20? It seems to me that's basically what AI researchers are trying to do, in a sense.

Answer 23

Passing the Turing test isn't a precise measure of the state of artificial intelligence science/technology. It would depend on the examiner & the topic of discussion.

A fundamental problem that isn't often brought up is that computers are deterministic - little more than state machines, actually - but we don't know yet whether humans are or not. If humans are capable of free choice, then we are by definition non-deterministic, which means there's a possibility that we can solve a wider class of problems than any Turing machine. If that's the case, deterministic computers will never achieve human-level intelligence.