A recent study revealed that the advanced language model GPT-4.5 passed the Turing test, traditionally used to gauge human-like intelligence. In a setup where participants engaged in conversations with both humans and AI, GPT-4.5 was perceived as human 73% of the time when it adopted a persona. This significantly surpassed random chance (50%), indicating a strong performance against both the Meta Llama 3.1-405B model and the earlier Eliza chatbot.
The Turing test, proposed by mathematician Alan Turing in 1950, involves a text-based conversation where an interrogator must identify the human and the machine participants. This latest experiment utilized nearly 300 participants and varied prompts for the AI, with “persona” prompts leading to higher success rates compared to “non-personal” prompts. Despite the impressive results, researchers, including Cameron Jones from UC San Diego, emphasized the complexity of assessing AI intelligence based on this test and cautioned against interpreting these results as definitive proof of human-like thought.
The findings raise concerns about AI’s capability to replace humans in various tasks, which could lead to job automation and other societal disruptions. The study underscores the evolving relationship between humans and technology, suggesting that as people interact more with AI, their ability to distinguish between human and machine might improve.
Source link
