Turing Test

The Turing test is an historical milestone goal in artificial intelligence, whereby a machine that passes is able to converse with someone and be indistinguishable from a human. In lay and pop science it is viewed by many as the single defining achievement of AI, but is seen as a distraction by many in the field. An example argument against the Turing test is that human communication isn't the only aspect of intelligence that exists, and reasoning and awareness without modern language could still indicate intelligence. An example argument in favour of it is that a system's intelligence is something that is measured in terms of its behaviour, and if its behaviour is indistinguishable from a known system defined as intelligent, then it is by definition measurably intelligent.

In my opinion, the Turing test isn't a test of intelligence at all, but of ability to mimic intelligence. For that reason alone I think it is more harmful than good, evolving AI work toward robust scripted responses instead of problem solving, cognition, thinking---the "hard" AI.

We're still in the infancy of AI, and it hasn't progressed as quickly as we once imagined (think Hal 9000 envisioned for the year 2001). Imagine if before air planes were invented, someone simply declared that the pinnacle of aircraft design would be if a person could fly between New York and Paris. It is arbitrary and does not directly evaluate design. This is like the Turing test. It has probably endured, because we haven't developed a truly intelligent system yet, and don't even know what it will end up looking like. A test of intelligence will evolve along with the technology, and we're just not there yet.

What might be a better test of artificial intelligence? I think that a more interesting milestone will be reached when an AI, instead of convincing a human that it is a person, is able to convince itself that it is. Surely a system that can think it is intelligent, is?

But then there is the problem that the easier it is to trick a system, the less indicative of intelligence it must be. We could not simply write a program that mimics the belief of introspective intelligence. And then again, how do we evaluate whether any system is mimicking belief, or truly believes? How do we do this when we do not even understand the process in humans? How can we be sure that our own thoughts are not just the product of patterns, of mimicking past thought processes? In that sense, mimicking a person well enough might be a sufficient test of intelligence. If a being (human or machine) convincingly argues that it is thinking or is conscious, and we're unable to probe it to tell if it is just saying so, thoughtlessly producing some programmed output, or is genuinely reasoning, how can we know?

The Turing test evaluates ability to display intelligent behaviour. Another important goal would be to solve a problem (but not a programmed one, or one of a class it is designed to solve. So I suppose the AI would need to figure out how to solve a new problem, and so evolve or rewrite itself, or at least build knowledge and ability). Another is to have self-awareness and feelings.

But the test... how do you test these things? Say an AI passes the Turing test and behaves like a human when probed. How then can one be convinced that it is thinking, having original thoughts, and not just producing them but... thinking... them... and how do we know that humans are really doing anything special anyway? We have our internal experience of thought... How can we prove that, or how do we internally know it's more than boring repetition of patterns?...

If we can't speak for certain about these things, I think we are not yet ready to define an ultimate test of what a true AI would be.