More and More People Struggle to Distinguish GPT-4 from Humans in Turing Test

TapTechNews June 17th News, the latest research shows that more and more people have difficulty distinguishing GPT-4 from humans in the Turing test.

According to TapTechNews, the 'Turing test' was proposed by computer scientist Alan Turing in 1950 and is also known as the 'Imitation Game'. The criterion is whether the machine can have a conversation like a human and make the other party mistake it for a real person.

Researchers at the University of California, San Diego recruited 500 participants and let them have a five-minute conversation with four 'conversationalists', which were a real person, the first-generation chatbot ELIZA in the 1960s, and GPT-3.5 and GPT-4 that drive the chatbot ChatGPT. Participants had to judge whether the other party was a human or a machine after the conversation.

The result of this test, published on the preprint website arXiv, showed that 54% of the participants mistook GPT-4 for a real person. In contrast, only 22% of people mistook the pre-programmed ELIZA for a real person, 50% for GPT-3.5, while the proportion of the human conversationalist being correctly identified was 67%.

Researchers pointed out that the Turing test is too simple, and the style and emotional factors of the conversation play a more important role in the test than traditional intelligent cognition.

The Institute of Electrical and Electronics Engineers (IEEE) artificial intelligence researcher Nell Watson said that pure intelligence does not represent everything. What really matters is to be smart enough to understand the situation, the skills of others and have the empathy to combine these elements. Ability is only a part of the value of AI, and understanding human values, preferences and boundaries is equally important, and it is these qualities that make AI a loyal and reliable steward in our lives.

Watson also mentioned that this study poses a challenge to human-computer interaction in the future, and people will be increasingly suspicious of the authenticity of the interaction object, especially when it comes to sensitive topics. At the same time, this study also highlights the great progress of artificial intelligence in the GPT era.