What Is The Turing Test And Will It Ever Be Beaten?
Is it possible for artificial intelligence to match human intelligence? It’s a tricky question that involves philosophy, psychology, computer science, and every topic in between. Whenever there’s talk about human-level machine intelligence, the Turing Test is never too far behind.
Earlier this month, Internet journalists exploded in a frenzy of excitement when a London-based computer program, named Eugene Goostman, seemingly passed the Turing Test . But what does that mean? Why are people so excited over it? And is it as noteworthy as it might first appear?
The Turing Test: What Is It?
Despite its name, the Turing Test is not a true test – at least, not in the common sense of the word. It’s actually more of a thought experiment that was put forth by Alan Turing, a highly influential mathematician (and one of the most famous geeks ) who formalized many of the concepts that led to the birth of computer science.
The Turing Test is a set of guidelines that are meant to determine whether or not a machine is indistinguishable from a human. It tries to answer the question, “Can machines think?” We don’t know the answer yet, but Turing believed it was possible.
The standard interpretation of the Turing Test is a kind of game:
- You are interrogating two people.
- Person A is a machine. Person B is a human.
- You can only communicate with them using text.
- By asking questions, determine which one is machine and which one is human.
At this time, the standard game length for the test is five minutes and the conventional criteria for passing the test requires the machine to fool at least 30% of all human interrogators. Turing predicted that any machine to do that could be “smart” enough to be labeled as a “thinking machine.”
Potential Drawbacks of the Turing Test
There are a number of problems with the Turing Test, however, and many arguments have popped up over the years regarding the design of the test and whether it truly assesses the thinking capacity of artificial intelligence.
The most obvious issue is the arbitrary nature of the testing criteria. Why is there a five-minute limit and why is the fooling rate of interrogators set at 30%? Why not ten minutes and 50%?
The truth is, those numbers were derived from Turing’s prediction about the future state of artificial intelligence . He never meant for them to be explicit thresholds. However, for now, those numbers are good enough as a target to reach.
Another major problem: a machine being indistinguishable from a human does not necessarily indicate intelligence. In other words, does the Turing Test actually test a machine’s ability to think for itself OR a machine’s ability to imitate human behavior?
It’s a subtle difference with huge implications. After all, a chatbot with enough lines of code could conceivably imitate human conversation without ever being truly intelligent. This brings up a subsequent question. Is external behavior enough to indicate internal thoughts?
The last major drawback to note is the lack of a control group. By definition, the Turing Test results are based on a group of interrogators, but not every interrogator in the group is equal. Though Turing does specify that the criteria are only relevant for “average interrogators,” it’s still unclear what “average” means.
Can The Turing Test Be Passed?
Let’s return to the earlier-mentioned Eugene Goostman program that supposedly passed the Turing Test. Does this now mean we can create human-like robots that are sentient and self-aware? Not quite.
One of the main criticisms against Eugene Goostman was the deceptive lowering of Turing Test criteria. The developers claimed the computer to be a 13-year-old boy who does not natively speak English and lives far enough away from modern society to be ignorant of topics such as geography, pop culture, etc.
In other words, they cheated. By framing Eugene Goostman in this context, interrogators did not have to hold the machine’s responses to a normal standard. After all, many modern chatbots are able to hold similar conversations. The difference with Eugene Goostman is that the narrative context surrounding the machine allowed the hiccups in conversation to be more believable.
So maybe Eugene Goostman didn’t properly pass the Turing Test. How about in the future? When will we see a machine pass it? Is it even possible?
Perhaps the better question is whether the Turing Test is still a valid measure of artificial intelligence. Many believe that the Test only encourages human imitation rather than true thinking intelligence. In fact, other AI tests have been designed in recent years that are more sophisticated and specific.
The Turing Test might be iconic, but maybe it’s time that we shelve it and move on.