For the first time ever, large language models have officially passed the standard three-party Turing test - with GPT-4.5 convincingly impersonating humans 73% of the time, even fooling interrogators more often than actual humans did.

In groundbreaking research from UC San Diego, scientists conducted rigorous three-party Turing tests where participants had 5-minute conversations with both a human and an AI simultaneously before guessing which was which.

Key findings:

→ GPT-4.5 (when prompted to adopt a humanlike persona) was judged to be human 73% of the time→ LLaMa-3.1-405B reached a 56% human judgment rate with the same persona prompt

→ Baseline models like ELIZA and GPT-4o achieved only 23% and 21% respectively

→ Appropriate prompting was crucial - models without specific persona instructions performed worse

Most participants didn't test for "traditional" intelligence (only 12% asked knowledge questions). Instead, they focused on social and emotional aspects of intelligence - conversation flow, linguistic style, and personality traits.

What's particularly fascinating is that interrogators were most accurate when using unusual strategies like saying strange things or attempting "jailbreaks."

When models lacked knowledge, this actually made them seem more human to interrogators!The question now isn't whether machines can pass as human, but how we'll redefine what makes us uniquely human in a world where machines can so effectively imitate us.