A recent preprint study found that OpenAI’s GPT-4.5 model was identified as human 73% of the time in a three-party Turing test when given a specific persona, significantly surpassing the random chance of 50%. The research also evaluated Meta’s LLaMa 3.1-405B and OpenAI’s GPT-4o models, with GPT-4o achieving only 21% identification success when prompted without personas. The study involved nearly 300 participants acting as interrogators and witnesses to assess the AI’s responses. 

Lead author Cameron Jones noted that the findings highlight that AI models, particularly when adopting personas, can effectively mimic human traits. However, he cautioned that the Turing test does not definitively prove human-like thinking in AI. Jones expressed concerns about the implications of LLMs potentially replacing humans in various interactions, leading to job automation and societal changes. The study also suggests that the ability to distinguish between humans and machines may improve as society becomes more accustomed to AI.

Source: Futurism