AI model passes turing test, fools humans 73% of the time

OpenAI’s GPT-4.5 has just surpassed the imitation game. In a recent preprint study, researchers discovered that when presented with a “persona,” GPT-4.5 was mistaken for a real human in 73% of cases, outperforming actual humans in certain rounds.

‍

This achievement surpasses the 50% benchmark set by Alan Turing’s renowned test for machine intelligence. However, when presented without a persona, GPT-4.5’s performance dropped to 36%. Meta’s LLaMa also demonstrated similar capabilities. Notably, OpenAI’s current GPT-4 achieved only 21% success. The study suggests that social engineering, job automation, and concerns about tech trust may soon become pressing issues.