“Any sufficiently advanced algorithm is indistinguishable from human intelligence.”
Alan Turing began his 1950 paper, “Computing Machinery and Intelligence,” with the proposal “to consider the question, ‘Can machines think?'” He went on to describe what is now commonly called the Turing Test as a way to characterize and identify artificial intelligence. With apologies to Arthur C. Clarke, I think the statement above is a concise way to express the motivation for and structure of the test.
To set the stage for the discussion, suppose that you are confronted with a robot that looks, walks, talks, and acts just like a human. Is the robot thinking? Does it have intelligence? Does it have the same sense of “me” that you or I have? Does it have a soul?
(I find these questions interesting in part because people often have very reflexive responses to them. If a robot isn’t human, then it might be able to play a good game of chess, or even perhaps fly an airplane, but it can’t possibly feel love, or pain, or anger… can it?)
One imagines that, to answer these questions, philosophers, scientists, and especially mathematicians would attempt to provide very precise, measurable, and most likely very complex and technical criteria that must be satisfied. Turing’s proposed test is surprisingly simple, however. It boils down to a challenge: hold a conversation with the robot and another real human subject… and try to determine which is which. If you guess incorrectly, then the robot has demonstrated behavior sufficient to be called “intelligent.”
(This is both more and less complex than the test as it is usually described. It is more complex in that the “conversation” is usually restricted to only text communication using a keyboard and monitor. It is less complex in that multiple conversations and “guesses” are needed to make a conclusion with some statistical certainty. As a thought experiment, I think the former difference is unnecessarily restrictive, and the latter is a technical point we can safely ignore for this discussion.)
This post was motivated by a 2006 paper by Stuart Shieber titled “Does the Turing Test Demonstrate Intelligence or Not?” It’s very short, just four pages, and does a very good job of summarizing not only the Turing Test itself, but also arguments from both sides of the fence regarding the test’s sufficiency as a means of demonstrating artificial intelligence. It also has some interesting mathematics (that’s how I stumbled on the paper in the first place), which if you are interested is described in more detail in Shieber’s longer follow-on paper, “The Turing Test as Interactive Proof.”
The papers make for interesting reading, but I think the critical point is made well upstream of the more sophisticated mathematics. When trying to argue whether a particular mass of flesh or of wires is– or is not– thinking, or even feeling, I think the test is as simple as Turing suggested: if you can’t tell the difference, then what is the difference?