Overview

• OpenAI's GPT-4.5 passed a Turing test with a 73% success rate when adopting a persona, surpassing the 50% threshold for passing and even outperforming human participants in some cases.
• The study, conducted by UC San Diego researchers, used a rigorous three-party Turing test involving nearly 300 participants and tested multiple AI models, including Meta's LLaMa-3.1 and the 1960s chatbot ELIZA.
• Persona prompts, which instructed the AI to adopt specific identities, significantly enhanced GPT-4.5's performance, compared to a much lower 36% success rate without such prompts.
• Meta's LLaMa-3.1 also passed the Turing test with a 56% success rate using persona prompts, though it lagged behind GPT-4.5 in human-like conversational performance.
• The results, published as a preprint and awaiting peer review, have sparked discussions about the limitations of the Turing test, the societal risks of advanced AI, and the need for new benchmarks to measure reasoning and ethical alignment.