On Turing Test and Artificial Intelligence

Alan Turing, the English computer scientist and mathematician proposed an experiment that becomes known by his name (the Turing test). In this test, a computer could be said “intelligent” if a human interrogator could not tell it apart, through conversation, from a human being.

 

Is this Test still valid today?

Is a machine that can pass the Turing test intelligent? Yet to answer this question we need to define what we mean by intelligence.

 

A thought Experiment

We can think about several “stupid algorithms” that could pass the test.

 

Suppose a machine that can store all the conversations that could exist. This machine can function only as a memory and respond by a sentence when a question is addressed by looking up the answer in its database. This machine can then, by a simple string-matching algorithm, pass the Turing test. Whenever the machine converse with a human it can only use its memory to interact.

 

Now to be fair, the matching algorithm and conversation manager can be fairly complicated, but the argument still hold. The idea is that the machine has no understanding of the current task and situation. It is a mere reflex system.

 

Can this machine be considered as intelligent? A similar argument was first published in 1980 by John Searle published “Minds, Brains and Programs” in the journal The Behavioral and Brain Sciences.

 

The real question behind the argument is: can simulated intelligence be the same as intelligence?

 

On the other hand, does the inner working of a machine really matter as long as it appears to be intelligent? Do we really have an idea about what is happening inside other humans brains (or ours in that matter) when we are interacting with them?  Aren’t we on a similar autopilot most of the time?

 

My opinion? it does matter. Our memory is merely a tool to spare the brain’s resources and to react instantly to known situations. We still can and do put “our intelligence” to work regularly.  Reproducing the less intelligent part of human cannot constitute true intelligence.

 

Clearly, Turing test cannot hold if we extend the definition of intelligence beyond conversation to the exhibition of the behavior and the flexibility of the humans in various situations.

 

While Turing test will be passed relatively easily soon, it is necessary to extend the intelligence test to interactions beyond conversations and include skills such as generic problem solving and abstraction. These skills involve competencies far greater and complicated than conversation.

 

This kind of tests lack of the simplicity of Turing test, but it is possible to define it in a more precise way.

 

 

 

The ultimate proof

Ultimately, once we put a test in place (any test), we can always build a machine that can pass it. And I guess we will still have the same philosophical discussions and questions about the real intelligence.

 

To accept machine intelligence as true intelligence, it is necessary that the processes that generates the intelligence be complicated enough to be opaque and not predictable.

 

Suppose that we build a machine that pass all the tests. Will we still be convinced that this machine has true intelligence if we discover that inner working is just a complex if then else state machine? My guess is that the answer is No.