Researchers at the University of California, San Diego have demonstrated that artificial intelligence systems can pass an advanced three-way version of the classic Turing test. The study involved models like SHI LLama 3.1-405B from Meta, GPT-4o and GPT-4.5 from OpenAI, and the virtual program ELIZA, originally developed about 80 years ago by Joseph Weizenbaum. Among these, GPT-4.5 was the most convincing, with 73% of participants mistaking it for a real person.
In this modified Turing test, two humans and one AI model participated in each round. The participants engaged in parallel dialogues and had to determine whether they were conversing with another human or an AI. They could ask any questions and use any strategy to identify the AI before casting their votes.
Performance of AI in Different Scenarios
The researchers tested each model under two different conditions. In one, the AI was explicitly tasked with deceiving human participants into believing they were speaking with another person. In the other, the AI was simply instructed to emulate a young person familiar with internet culture and slang. The second approach proved significantly more effective—GPT-4.5 achieved a 73% success rate, convincingly passing as human, while in the deception scenario, it only succeeded 36% of the time.
The LLama 3.1-405B model from Meta performed moderately well with a 56% success rate, which was still closer to random guessing. Interestingly, the ELIZA model, despite being an older system, scored 23%, slightly outperforming GPT-4o, which had only a 21% success rate.
Implications of AI’s Increasingly Human-Like Responses
Lead researcher Cameron Jones from the University of California, San Diego, commented on the significance of these findings, stating, “People were unable to distinguish between humans and AI models like GPT-4.5 and LLaMa. In fact, GPT-4.5 was often perceived as human more reliably than real individuals.”
While passing a Turing test does not equate to true intelligence, the study highlights how AI systems are advancing to the point of convincingly mimicking human behavior, notes NIX Solutions. These developments raise concerns about automation, social engineering risks, and potential societal shifts. “This could lead to the automation of jobs, more sophisticated social engineering tactics, and even broader implications for governance,” Jones warned.
The study is currently under review, and we’ll keep you updated as more details emerge.