GPT 4.5 fools people into thinking it was human, passes Turing test with 73% success rate

Chatgpt

In a groundbreaking study, researchers at the University of California, San Diego, found that some AI models are now so good at mimicking human conversation, they can actually pass as people.

The researchers used something called the Turing Test. It’s named after British mathematician Alan Turing, who, back in 1950, came up with a simple question: if you can’t tell whether you're chatting with a human or a machine, does it really matter? In the modern version of the test, a person chats separately with both a human and a machine and then has to guess which one is the real deal.

In this study, three AI models—OpenAI’s GPT-4.5, Meta’s Llama 3.1 405B, and OpenAI’s GPT-4o—were put to the test. Human participants had five-minute conversations with both a real person and an AI without knowing who was who. Then, they had to pick who they thought was the human.

Here’s where things get crazy: GPT-4.5, when told to act like a human using what they called a “PERSONA” prompt, fooled people 73% of the time. That means people actually picked the AI more often than the real human! Llama 3.1 405B also did well, fooling people 56% of the time. GPT-4o, without any personality boost, only tricked people 21% of the time.

Most of the conversations were about day-to-day stuff—what the person did that morning, what they thought about certain topics, even how they felt. Basically, small talk. And the AI held its own, blending in surprisingly well.

Some experts say this kind of tech could easily slide into roles that rely on quick conversations—think customer service or even casual online chats. Others are a bit more cautious, wondering what happens when AI starts replacing real people in social spaces. Friends, coworkers, maybe even romantic partners?

One thing’s for sure: AI is no longer just answering our questions—it’s joining our conversations. Just don't be surprised if your next favorite pen pal turns out to be silicon-based.