Oct 28th 2020

The Turing Test Aims to Understand Advancing Computers


If you could go back in time to 1950 and have Alan Turing speak to a chatbot with a smartphone, he might declare a winner of his test to determine if a computer was “intelligent.” With the invention of the mobile phone 23 years away and the voice-powered virtual assistant inside smartphones even further in the future, no one in Turing’s time would have imagined that computers could be as advanced as the ones we take for granted today — or that they could also fit inside a pocket.

At its essence, the Turing Test challenges computer scientists to create a program with a conversational ability that a judge can’t distinguish from a human. Some scientists contend that taken at face value, the test has already been aced by AI-supported chatbots. Surely, at some point, you’ve wondered if the “representative” answering your question on a company’s website was either a well-programmed chatbot or a human being.

But just because we can be occasionally tricked into thinking that a chatbot is human doesn’t mean the 70-year-old Turing Test is in any way dated, that it could apply only to an era when the expectations of computers were lower, when large mainframes didn’t actually speak and the idea of a chatbot would have seemed like science fiction. The test, in fact, is even more important today as artificial intelligence gains new abilities, as well as prompting us to consider what it means to be human. The ability to understand artificial intelligence can give us insights into our strengths and weaknesses as humans.

“For artificial intelligence to fully pass the Turing Test, it would have to understand what it means to be human,” said Bruce Swett, the chief AI architect at Northrop Grumman. “We have all sorts of relationships that define us — with children, parents, friends, colleagues, groups, and organizations. There is a matrix of social interconnectedness that we intuitively know. No AI model understands that yet. That will take a long time.”

Turing and Computing

A polymath if there ever was one, Alan Turing was a mathematician, computer scientist, biologist and even a philosopher. He was also a cryptanalyst, and his first claim to fame was cracking the Enigma code that German armed forces used to obscure messages of military strategy during World War II. By breaking what the Germans thought was an unbreakable code, Allied forces were able to avoid Nazi submarines that had sought to destroy military and commercial cargo ships.

But even before the war, Turing believed in the power of computing. He invented the Universal Turing Machine, which isn’t a machine but rather a formula. According to Andrew Hodges, who wrote the biography “Alan Turing: The Enigma,” this formula allows a machine to tackle “any well-defined task by being supplied with the appropriate program.” This was a profound concept. It was the precursor to the type of computing we use and rely on today.

As Hodges wrote: The “Universal Turing Machine naturally exploits what was later seen as the ‘stored program’ concept essential to the modern computer: it embodies the crucial twentieth-century insight that symbols representing instructions are no different in kind from symbols representing numbers. But computers, in this modern sense, didn’t exist in 1936. Turing created these concepts out of his mathematical imagination.”

The Turing Test Expanded

Recognizing the potential of computing, Turing wondered if machines could ever possess intelligence on a par with, or even surpass that of, humans — so in 1950 he created his famous challenge. Known by its creator as “The Imitation Game,” the Turing Test tries to discern whether a computer can mimic human intelligence.

In the Turing Test, a judge exchanges texts with participants using computers, but can’t see the participants. The participants are either a human or a computer program that attempts to trick the judge into believing it’s human through its ability to communicate intelligently via text.

It’s widely believed that no computing program has passed the test, but one came close. In 2014, a program convinced 10 out of 30 judges of an AI competition at the Royal Society in London that it was a 13-year-old Ukrainian by the name of Eugene Goostman. At the time, many scientists disputed that “Eugene” had actually passed the test, although one of the program’s creators found a silver lining in the results, telling The Guardian, “I think we passed ‘a’ Turing test, but I don’t know if it’s ‘the’ Turing test.”

Turing believed that a computer would win his Imitation Game by 2000, said Bruce Swett, who is a Northrop Grumman Fellow. Informally, AI-powered chatbots are, in fact, beating the Turing challenge now, as they often convince people that the chatbot is another human, Swett said. However, occasionally convincing a person that an AI program is human is not the same as passing a formal Turing Test.

As it was conceived, the Turing Test is limited to one sensory “channel,” a written conversation that’s separated by a metaphorical curtain, Swett said. The judge doesn’t get visual or auditory input, which would make it easier to decide whether the message came from a human or an AI. But now, as AI capabilities develop, Turing’s test should be expanded conceptually, he said.

For example, we’re now interacting with AI through sight and sound instead of just text, a prompt to expand the Turing Test and include more of the five senses in distinguishing what’s human and what isn’t. “AI-generated deepfake capabilities can convincingly create a lifelike video of someone doing something – like giving a speech – that they never did,” Swett said. These additional sensory channels will need to be included in future versions of the Turing Test to fully challenge emerging AI capabilities.

Finding Ourselves in the Computers

A computer algorithm that passes an enhanced Turing Test would have to interact with the world in a way that humans do, Swett said. “We’re amazingly good at making sense of the environment around us. We use our senses and really have a rich, integrated understanding of the environment, identifying people, animals, objects, and locations. That provides a context to understand the world around us by using our sensory experience.”

For instance, we understand cats by observing them, touching them, listening to them hiss, Swett said. “Our brain pulls together all that detailed sensory information and brings it into any discussion about cats, including predictions about how we think a cat will behave. Similarly from a motor context, when we see someone else kicking a ball, a part of our brain actually warms up the process that prepares us to kick a ball. We understand our world from our sensory and physical interactions with it.”

For AI to not only pass an enhanced Turing Test but to function on a level with humans, such AI programs would need to be, in some way, embodied like humans. This would allow the AI to interact with the environment, collecting sensory and physical interactions. “To move from current, narrow artificial intelligence to artificial general intelligence, the AI would have to understand what it means to be human,” Swett said. From there, it could fully understand what makes a cat a cat, or what it means to kick a ball.

Until then, we can observe with amazement as computing continues to advance, and equally find remarkable that a polymath who didn’t own an iPhone saw all of this coming.

Check out Northrop Grumman career opportunities to see how you can participate in this fascinating time of discovery in science, technology, and engineering.