GPT-4.5 Achieves Human-Level Performance in Turing Test, Fools 73% of Participants

A recent study indicates that OpenAI's GPT-4.5 model has demonstrated human-level performance in a three-party Turing Test. The research, conducted by the University of California, San Diego, found that GPT-4.5 was identified as human 73% of the time when prompted to adopt a specific persona. This significantly exceeds the 50% baseline, suggesting that participants often mistook the AI for a human. The study involved nearly 300 participants who engaged in text-based conversations with both a human and an AI model. The interrogators' task was to identify the human participant. The AI models were evaluated under two conditions: with a minimal instruction prompt and with an enhanced persona prompt that guided the AI to adopt a specific human-like demeanor. The research also evaluated Meta's Llama 3.1-405B model, OpenAI's GPT-4o model, and an early chatbot known as ELIZA. The results indicated that GPT-4.5, when using the persona prompt, achieved a win rate of 73%. Llama 3.1-405B with the persona prompt attained a win rate of around 56%, whereas GPT-4o under no-persona conditions only reached a win rate of 21%. This suggests that the persona prompt significantly improved the AI's ability to mimic human conversation. The study highlights the advancements in AI's ability to simulate human interaction and raises questions about the implications of such technology.