The Turing test is an experiment to determine whether a computer program possesses artificial intelligence (AI). Here's everything you need to know about the Turing Test, including its definition, history, limitations, variations, and major applications.
The purpose of the Turing test is to see if a person can tell the difference between another person and a computer program.
Initial testing consisted of two human subjects and an autonomous computer. A person (the questioner) sits in a room with a keyboard and monitor. The other person (the responder) is in a separate room with a keyboard and monitor. The computer itself is located in another room.
Human test takers enter a series of predetermined questions, which are received by both the respondent and the computer. Human respondents and computers provide their answers, and test takers must determine which responses come from humans.
A computer is considered artificially intelligent if a test taker cannot tell the difference between a human and a computer more than half the time after repeating the test.
Can a human being differentiate between another human being and a computer program?
The Turing test was invented by British computer scientist Alan Turing, who first proposed the idea in his 1950 paper "Computing Machines and Intelligence."
Turing's idea was based on a similar experiment called the "Imitation Game," involving two men and a woman. One of the men had to determine the gender of the other two subjects based solely on their responses to the same questions.
The problem, of course, is replacing one of the subjects with computers.
Turing's work inspired artificial intelligence chatbots such as ChatGPT. It and other products have passed the Turing test in controlled settings.
In 1966, Joseph Weizenbaum developed a chatbot called ELIZA that met the basic conditions of the Turing test: it could generate sentences that were indistinguishable to the user, written by a computer program .
The Loebner Prize is awarded annually from 1991 to 2020 to the artificial intelligence that performs best in the Turing Test.
In popular culture, "passing the Turing test" means that artificial intelligence can compete with humans.
In 2014, a chatbot named Eugene Goostman won the Turing Test competition organized by the University of Reading to commemorate the 60th anniversary of Alan Turing's death.
This chatbot with the personality of a teenage boy convinced a third of the judges that it was human.
For example, in a live demonstration in 2018, AI engineers used Google Duplex to make reservations over the phone without the human on the other end realizing they were talking to a computer. Some media reported that Google Duplex had passed the Turing test, but no controlled test was actually conducted.
In the early days of artificial intelligence research, Turing test questions had to be limited to specific topics. They must also be limited to answering "yes" or "no." If the question is too open-ended, test subjects can quickly tell whether the answer came from a computer.
Because the traditional Turing test is limited to written text and must be administered in a controlled environment, it is a very narrow measure of intelligence.
The criteria for human-like intelligence have changed over the years, and philosophical debates over the meaning of "intelligence" continue today.
Instead of trying to make users believe they are talking to another human, AI engineers are turning their attention to finding practical applications and improving interactions between humans and AI.
Over the years, the concepts behind the Turing Test have been expanded and adapted, with many variations developed, including:
- Reverse Turing Test : Humans must convince computers that they are human. Many websites use reverse Turing tests (in the form of CAPTCHAs) to detect bots.
- Minimum Intelligent Signal Test (MIST) : A traditional Turing test that uses only Boolean or binary responses (i.e., true/false or yes/no).
- Full Turing Test : In addition to verbal reasoning, questions test perceptual skills (such as the ability to identify and manipulate objects). Such tests are used to develop artificial intelligence-powered robots.
- Lovelace Test 2.0 : Tests the ability of artificial intelligence to create art based on text prompts.
There is no standardized set of Turing test questions, in part because programmers can design their software to pass it. Instead, the judge asks his or her own questions. Some possible questions involve childhood memories as well as questions requiring more creative, lateral thinking (e.g., "Don't use adjectives to describe yourself").
The Turing test is useful for measuring progress in natural language artificial intelligence. What passes—or even comes close to passing—is considered particularly advanced. The test also provides data that developers can use to improve their programs by showing where they struggle.
