“Turing is one of my all-time heroes,” Demis Hassabis, co-founder of DeepMind, told Lex Fridman recently. However, he also immediately says that the original paper published by Alan Turing in 1950 did not address it to be a formal test. For years, the use of the Turing Test to check the viability of a machine has been like a pandora box.
There is no denial that the paper published by Turing led to a major revolution in computing and the field of machine learning. However, not everyone agrees that it is a valid test for an AI system and Hassabis says the paper is actually written like a thought experiment. Let us take a look at the Turing test and why it manages to split opinion across the board.
What is the Turing Test?
One of the end goals of artificial intelligence (AI) is to build a machine that is capable of thinking like a human being. In other words, a machine that mimics the human brain and the Turing Test acts as a method of inquiry or a way to check whether it is possible for a computer to think like humans.
Named after Alan Turing, an English computer scientist, cryptanalyst, mathematician, and theoretical biologist, the test proposes whether a computer can be said to possess AI capabilities that can mimic human responses under specific conditions.
The original Turing Test, which is a twist on “The Imitation Game”, requires three terminals or stakeholders and each of which is separated from the other two. One of the terminals is operated by a computer while the other two are operated by humans.
How does the Turing Test work?
A computer is tested to possess artificial intelligence by using all the three terminals defined in the Turing test. During the test, one of the humans act as questioner while the other human and computer act as respondents. The human acting as a questioner tests the respondents within a specific area.
The test also uses a specified format and context. After the questioner asks a set number of questions, this human is asked to decide which respondent is human and which respondent is a computer.
This test gets repeated multiple times before reaching a conclusion. If the human questioner manages to make the correct distinction between another human and machine in half of the test runs, the computer is considered to have artificial intelligence. The reasoning here being that the questioner is regarding the computer “as human” and matching responses with the human respondent.
What are the limitations of the Turing Test?
The biggest limitation, according to scientists and AI practitioners, in the past few years has been over the questioning and its narrow scope. Since the questions are restricted to a specified format and context, there have been vocal criticisms of this approach.
For one, Hassabis believes that this approach should not be termed as a scientific way to determine whether a computer possesses artificial intelligence. ELIZA, one of the first programs to pass the Turing Test, was found manipulating symbols it did not understand fully.
There have been instances where a computer has been found to score higher if the questions where objective type and answers were restricted to “Yes” or “No”. If the questions were open-ended and required conversational answers then computers struggled to successfully fool the questioner.
Thus, the question of whether a computer can pass the Turing Test has become irrelevant for a number of AI researchers. Companies like DeepMind are even prioritising human-machine interaction and trying to make it intuitive.
What is the importance of the Turing Test?
“A computer would deserve to be called intelligent if it could deceive humans into believing that it was human,” Turing wrote in a 1950 paper describing the now-famous Turing Test.
At the time when it was first defined, the idea of the test was meant to test machines that came before it. However, we have now seen how AI companies try to validate their computers or programs using the Turing test.
In many ways, the importance or relevance of the Turing test also depends on the humans. Imagine a human pretending to be a computer online and a Turing test being successfully able to identify this person as a human. Humans are increasingly interacting with machines and some of them are really good.
At a I/O developer conference, Google demonstrated its Duplex system capable of booking appointments on behalf of its user. Regardless of whether you consider the Turing Test as an important test for AI, it does act as the first step in understanding or identifying the progress of a computer.
What are the variations of the Turing Test?
In order to make the Turing test more relevant for modern AI systems, there have been a number of variations of the original test. All these variations are modern approaches meant to better detect humans and machines. They also continue to evolve in order to stay relevant. Here is a look at some of these variations:
- The Reverse Turing Test: This test reverses the original script by having a human trick a computer into believing that it is not interrogating or questioning a human.
- The Total Turing Test: This modified test allows the questioner to also test perceptual abilities as well as the ability to manipulate objects.
- The Minimum Intelligent Signal Test: This approach allows asking the test subjects only binary questions, which means the answers are either true or false or Yes or No.
- The Marcus Test: The test subjects view media and are required to answer questions about the content consumed.
- The Lovelace Test 2.0: To see whether a computer has artificial intelligence, this test makes the computer create an art and examines its ability to do so.
- Winograd Schema Challenge: This test also asks multiple-choice questions to the respondent but limits the questions to a specific format.
How is the Turing Test used today?
One can argue that some of the variations mentioned above are more relevant to testing machine intelligence of computers today. However, we still see the Turing Test being used in its original format. The most popular use of the Turing test is seen in the form of the annual Loebner Prize awarded since 1990.
This award is given to the most human-like computer selected by a panel of judges. The Loebner Prize follows the standard rules of the Turing Test and some critics call it publicity.
In order to mark the 60th anniversary of Turing’s death in 2014, the University of Reading organised a competition. In this competition, a chatbot called Eugene Goostman simulated a 13-year-old and passed the Turing Test. It managed to fool the judges 33 per cent of the time. This victory has been criticised a lot with some arguing that there weren’t enough judges.
Google Duplex, which debuted at search giant’s annual developer conference in 2018, successfully booked an appointment with a hairdresser. The receptionist could not decipher that they were not interacting with a real human. Although it did not use the original form of Turing test, the Duplex from Google can be described as a modern AI system to pass the Turing test.
Is the Turing test even relevant?
Despite all the criticism surrounding the test, the Turing Test has proven not only to be relevant but also widely used right now. The test mainly stands as a philosophical starting point for researching the AI capabilities of a computer. The Turing Test should be described as foundational for testing advancement in AI.