Artificial intelligence has taken a major step forward with GPT-4, a recent model that appears to have pushed the boundaries of simulated human interaction. A recent study found that GPT-4 was indeed successful 54% of the time in a revisited Turing test, outperforming its predecessors like GPT-3.5 and the 1960s ELIZA program which respectively scored 50% and 22%.
The Turing Test: evaluating artificial intelligence
Developed by OpenAI, GPT-4 uses advanced neural network architecture to process and generate text. Its ability to understand and respond contextually to questions and dialogue allows it to surpass its predecessors and come closer to authentic human interaction. This performance is based on years of learning from large language datasets, allowing GPT-4 to generate responses that feel natural and relevant to users.
In recent work, researchers have passed the Turing test to GPT-4. Proposed by Alan Turing in 1950, this test is a classic test bed in the field of artificial intelligence. Its objective is to determine whether a machine can simulate a human well enough so that this one cannot distinguish whether the interlocutor is a machine or a human being through a text conversation.
The study therefore aimed to determine the extent to which GPT-4 could fool participants into believing that they were conversing with a human being rather than with artificial intelligence. To do this, the researchers organized sessions where 500 participants were invited to engage in text conversations with four different interlocutors: a human being, the program ELIZA (a 1960s system with pre-programmed responses), GPT-3.5 and GPT-4. Each conversation lasted five minutes, after which participants had to guess whether they were speaking with a human or an AI.
Results and observations
GPT-4 succeeded in convincing the participants in 54% of cases that they were talking to a human being. GPT-3.5 reportedly scored 50%, while ELIZA was judged human only 22% of the time, highlighting the stark difference between the capabilities of modern AI models and older approaches.
The ability of GPT-4 to understand the context of conversations was crucial to his success in the Turing Test. The model is indeed capable of synthesizing responses that take into account the previous context of the conversation, the linguistic nuances and the subtleties of the questions asked, which helps to create an illusion of authentic human interaction.
The study also poses important questions regarding the evolution of artificial intelligence and its potential applications. Although GPT-4 has shown impressive capabilities, its use also raises ethical concerns, particularly regarding the transparency of human-computer interaction and the socio-economic implications of widespread use of such technologies.
