AI Beats Turing Test: GPT-4.5 Fooled 73% of Human Judges

An AI model has been found to pass the famous Turing test.

In a new study, GPT-4.5 was mistaken for a human 73% of the time. This happened when the model was told to act like a real person. The test had three people: one human, one AI, and one judge. The judge could not see who was who. The task was to guess which one was the AI.

The result showed that the AI seemed more human than the real humans. Even Meta’s LLaMa 3.1-405B and OpenAI’s GPT-4o were tested. A very old chatbot, ELIZA, was also part of the test.

GPT-4.5 did better when it used a “persona” prompt. Without this, it was judged human only 36% of the time. GPT-4o scored just 21%, while ELIZA got 23%.

The Turing test was created by Alan Turing in 1950. He said a machine might be called intelligent if a person could not tell it apart from a human in a text chat.

This study was done online. Nearly 300 people took part. Each round had one judge and two people being tested—one human and one chatbot.

In one version of the test, the AI was told to pretend to be someone who knew internet culture. This helped it sound more human. With no such prompt, the AI did not perform as well.

Experts say that passing the Turing test does not prove that AI can think. The test was more of a thought experiment than a real test. AI is trained to sound like us, even when it does not understand the topic.

This is why some researchers think the Turing test may no longer be useful. AIs like GPT-4.5 are very good at copying human speech. So it is harder to say if they really "think."

Still, the study shows that AIs might replace people in short chats. This could change jobs, lead to scams, and affect daily life.

The Turing test also shows how humans see machines. As people get used to AI, they might get better at spotting it too.

📌 Source : futurism