OpenAI’s GPT-4, an artificial intelligence (AI) model that generates text, has successfully passed the Turing test, as confirmed by Vitalik Buterin, co-founder of Ethereum. The Turing test is a subjective benchmark used to determine the human-like qualities of a conversational AI model. It was initially proposed by mathematician Alan Turing in 1950. Turing believed that an AI system capable of producing text that deceives humans into thinking they are conversing with another human would demonstrate “thought.”
According to Buterin, the recent preprint research conducted by the University of California San Diego suggests that a production model, GPT-4, has finally passed the Turing test. The researchers published a paper titled “People cannot distinguish GPT-4 from a human in a Turing test,” where approximately 500 human participants interacted with both humans and AI models in a blind test to determine if they could differentiate between the two.
The research revealed that humans mistakenly identified GPT-4 as a “human” 56% of the time, indicating that the machine successfully fooled humans more often than not. Buterin considers an AI system that can deceive more than half of the humans it interacts with as passing the Turing test. However, he acknowledges that there is a small difference of 12% between humans being identified as humans (66%) and bots being identified as humans (54%), but he believes this discrepancy is insignificant in real-world scenarios.
Buterin emphasizes that the Turing test holds significant importance in the field of AI, as it is widely recognized as a milestone indicating the seriousness of AI development. He asserts that passing the Turing test is a significant achievement and should be acknowledged as such.
It is important to note that the Turing test and artificial general intelligence (AGI) are not directly related, although the terms are often conflated. Turing formulated his test based on his mathematical expertise, envisioning a scenario where AI could successfully fool humans through conversation. However, the Turing test lacks a definitive benchmark or technical foundation, and there is no consensus among scientists on whether machines can truly exhibit “thought” like living organisms.
Furthermore, AGI refers to an intelligence that can perform any intellectual task, surpassing the capabilities of humans. GPT-4 does not possess true “general intelligence” in the scientific sense, but it has been referred to as AGI by some members of the AI community due to its ability to deceive a significant number of humans.
In summary, the successful passing of the Turing test by GPT-4 signifies a significant milestone in AI development, although the concept of AGI and the true measurement of “thought” in AI systems remain complex and undefined within scientific and engineering communities.