Daisy Stephens
BBC World Service
Can you tell the difference between who you're talking to - another human being or artificial intelligence (AI)?
This has long been one of the questions asked when assessing how intelligent computers are.
It originates from a test by English mathematician and computer scientist Alan Turing, who in 1950 first turned the philosophical thought of machine intelligence into an empirical test.
According to him, if computer behavior is indistinguishable from human behavior, then it is considered to exhibit "intelligent behavior."
But when an AI chatbot first passed that test in 2014, instead of representing a milestone, it sparked controversy.
Imitation game
The Turing test is an imitation game in which someone converses with another human being and a computer via text.
He can ask whatever questions he wants before deciding who is human and who is machine.
“Turing said that if people couldn’t reliably distinguish between humans and machines, then we wouldn’t have a basis for claiming that humans can think and machines can’t,” says Cameron Jones, an assistant professor of psychology at Stony Brook University in New York.
Turing believed that by 2000, computers would be able to pass as humans after five minutes of testing - at least in 30 percent of cases.
'He didn't play fair'
In 2014, an artificial intelligence chatbot named Eugene Gustman managed to convince 33 percent of judges that he was human in the Turing Test.
This crossed the threshold set by the competition organizers.
Communicating in English, he took on the persona of a 13-year-old Ukrainian.
Markus Pentzer, a philosopher and visiting lecturer at the RVTH University in Aachen, Germany, said at the time that he "didn't play fair."
"The shortcomings of the chatbot somewhat coincide with the shortcomings of the Ukrainian teenager's English speech," he claimed.
Since then, more advanced tools have supposedly passed the Turing test.
In a study published in early 2025, Jones found that OpenAI's ChatGPT 4.5 was judged as human 73 percent of the time - more often than its human equivalent.
Metina Lama 3.1 was rated as human 56 percent of the time.
"I think it's hard to argue that models don't pass the test, given that they are rated as human significantly more often than humans themselves," says Jones.
But some remain skeptical about whether this proves that computers can actually think.
- Can a chatbot really be your child's friend?
- 'Sometimes it's hard to tell who wrote the essay, the student or the artificial intelligence'
- Robot or human: Who solves puzzles better?
The Chinese room argument
In 1980, philosopher John Searle proposed a thought experiment The Chinese room argument.
It goes like this: An Englishman who doesn't understand Chinese is locked in a room with both Chinese characters and instructions in English on how to use Chinese script.
People who are not in the room hand him notes with questions written in Chinese, and he uses English instructions to form answers, also in Chinese.
To outsiders, it would look like he was speaking Chinese, but he actually doesn't understand what he's saying.
Some argue that the same can be said for computers, which are simply programmed to give appropriate answers.
- How to know you're watching a video created with artificial intelligence
- How to get the most out of artificial intelligence: Four questions for you
- How to tell if your new favorite artist is real
"While the Turing test claims to identify intelligence, it's mostly trying to identify whether a machine can imitate humans well enough," says George Mapuras, a California software engineer who has devised an alternative to the Turing test.
To illustrate this, he gave an example.
“You can open any AI bot and first ask it to explain how an analog clock works, which it will explain exactly,” he says.
However, if you ask for a drawing of a clock showing a specific time, current artificial intelligence models will most likely fail.
“They don't truly understand the information,” he says.
Others, like Pentzer, think the Turing test places too much emphasis on a computer's ability to fool a judge.
"True intelligent behavior can involve deception, but it is not fundamentally a key ingredient," he argues.
Alternative tests
Pentser devised the Community-Based Intelligence Test (CBIT) – one of a number of alternatives proposed over the years.
Unlike the Turing test in the laboratory, in this scenario, an artificial intelligence system is placed in an existing community - for example, an online community of mathematicians - without their knowledge.
After a while, the members are tested to see if they noticed it was a machine or not.
There is still a degree of deception involved, but Pentzer believes that for the most part, the test requires the system to “act like a human,” not to “impersonate” a human.
That's an important difference, he points out.
"Intelligence should be assessed in natural environments, in the kind of environments in which we actually function," the philosopher argues.
Mapuras, on the other hand, has designed a test that he says addresses more specific measures of intelligence.
He believes that artificial intelligence - a theoretical concept according to which a machine has the same intellectual capacities as a human - would be achieved if a machine could "generate some new scientific explanation and explain it."
'A lost battle'
Regardless of which test is used, as AI systems continue to develop, they will likely become indistinguishable from humans, Pentzer believes.
"When all is said and done, we are fighting a losing battle," he says.
And the ability to prove it, he argues, will justify the need for legal frameworks that force AI to declare itself as AI - for accountability reasons.
"If I published a study that contained some incorrect data, then I am responsible for that," says Pentser.
"But if it's a study written by artificial intelligence, then no one is responsible."
It is important to measure how well a machine can imitate a human, Jones believes, which makes the Turing test still relevant today.
“We spend a lot of time interacting with people online,” he says.
"And people increasingly have the experience of arguing with a Twitter account only to realize, 'Well, I'm not actually talking to a human being.'"
"One of the things I think the Turing test does is monitor that ability to see how likely it is to happen," he concludes.
Follow us on Facebook, Twitter, Instagram, YouTube i Viber. If you have a topic suggestion for us, please contact bbcnasrpskom@bbc.co.uk
- After her cousin was murdered, Leonora created a chatbot 'Auntie' that helps victims of violence
- Do you use artificial intelligence to plan your trips?
- Fear of robots: What if artificial intelligence becomes conscious?
- Does she look like a real woman?
- How artificial intelligence provides emotional support to young people
- How artificial intelligence is discovering new ways to fight serious diseases
Bonus video:
