Las Cadboods of Artificial intelligence (AI) as the Chadgiptd they are designed to reflect as true as possible to human speech In order to improve user experience.
But as AI is growing increasingly, it is difficult to distinguish these models from real people. And now, Science has proved it.
Scientists University of California (UCSD) in San DiegoAmerica, reveal In a study Those two Cadboods Also used in an important milestone used: GPD, which runs it Chadds Opanai, as he calls, is behind Cole AI on WhatsApp and Facebook, and has passed the popular touring test.
Designed by mathematician and comprehension of British Indexes of World War II Alan Touring In 1950, the Turing or ‘Hairy Game’ is a standard test designed to test the intelligence of a machine: an AI test passes when he is a human being Cannot be exactly distinguished Between another man’s response and the AI response.
“There are results that are the first empirical evidence that any synthetic system executes a stable three -part test,” says UCSD scientists. “If investigators cannot distinguish the credibility between a man and machine, sElectr says the machine has passed the test“They are adding.
This is the test
Researchers They used four models From AI: GPT -4.5 (launched in February), the previous reconstruction GPT -4 O, Meta, Lama and the 1960s chat project is called Elisa.
The first three ‘Big Language Models’ (LLM), deep learning mechanisms to identify and create text based on knowledge derived from mass data packages.
Experts enrolled 126 university students from the University of California in San Diego and 158 from the online database. There were participants Five -MINITED Internet Conversations At the same time, they had to decide what they thought was a man with another human participant and one in AI, but they did not know.
When asked to follow a human personality, it was considered GBD -4.5 This is the man’s 73% of the timeMore often than the real human participant.
Such a high percentage suggests that people are more efficient than people’s chance to determine whether GPT -4.5 is a man or machine. In the meantime, call -3.1, when he was asked to follow a human personality, 56% of time was considered human.
The team said that “is not substantially more or less than the men who compared them, but still Think of it as recognized.
Finally, reference models (Elisa and GPD -4 O) have reached significantly low success random rates: 23% y 21% Respectively.
Researchers tried to provide more basic instructions to samples and told them to adopt them without detailed instructions A personality like man.
As expected, AI models had performance In this position is significantly bad, This exemplifies the importance of implementing first Cadboods.
The group suggests that their new study has been published as a forecast “Solid Certificates” Opanai and Meta Pods have passed the touring test.
“This should be assessed as one of the many sources of the intelligence department shown by LLMS,” the main writer said, Cameron JonesIn a thread of X.
Jones agreed that when AI was working better Were told in advance to follow a manBut this does not mean that the GPT -4.5 and the lamas did not pass the touring test.
“Without any warning, for trivial reasons (as they admit to AI) will fail LLMS They can easily be adjusted to behave because they are asked, So I think it is fair to say that LLM agrees, “Jones emphasizes.
