Overview

• Researchers at UC San Diego conducted a Turing test with nearly 300 participants, where GPT-4.5 outperformed humans and other AI models when adopting personas.
• The study found GPT-4.5 achieved a 73% success rate in being identified as human, surpassing the 50% threshold historically considered a 'pass.'
• Without persona prompts, GPT-4.5's success rate dropped to 36%, highlighting the significant role of contextual instructions in AI performance.
• Other tested models, including Meta's LLaMA 3.1-405B and OpenAI's GPT-4o, performed worse, with GPT-4o achieving just a 21% success rate.
• The findings, published on a preprint server and awaiting peer review, have sparked discussions about ethical concerns, including job automation and AI-driven misinformation.