A new study shows that GPT-4.5 can trick people into thinking it’s human, ZDNET reports. Alan Turing created the Turing Test to check if machines act like humans. In the test, a human judge chats with two witnesses. One witness is human, and one is a computer. The judge guesses which is human based on text messages. Researchers at the University of California at San Diego tested GPT-4.5. They found it fools judges 73% of the time. This beats humans, who convince judges less often.
The led by Cameron Jones and Benjamin Bergen and published in arXiv with the title "Large Language Models Pass the Turing Test," used students as judges and witnesses. GPT-4.5 used a special prompt called PERSONA. A prompt is instructions given to AI. PERSONA told GPT-4.5 to act like a young, introverted person who knows internet slang. This made it seem more human. Other AIs, like GPT-4o and an old program called ELIZA, scored lower. ELIZA, a simple chatbot from the 1960s, fooled judges 23% of the time. Judges thought ELIZA’s odd replies seemed human.
Test measures humanlikeness, not intelligence
Melanie Mitchell, an AI expert, says the Turing Test doesn’t prove intelligence. Intelligence is the ability to think and learn like humans. The test checks if AI seems human, not if it thinks deeply. GPT-4.5’s success shows it adapts well to prompts. It mimics human chat styles convincingly. But judges often focus on sociability. For example, judges liked ELIZA because it seemed sarcastic, not smart.
The researchers say the test reveals human assumptions. People expect AI to act a certain way. If AI breaks those expectations, people think it’s human. The study suggests intelligence isn’t enough to pass. AI must mimic human habits, like casual chat. Jones and Bergen propose new test designs. For example, AI experts as judges might spot differences better. Adding rewards could make judges think harder.
The Turing Test may need updates. It struggles to measure complex intelligence. Researchers now involve humans to evaluate AI more closely. Some wonder if machines will need to judge AI intelligence in the future. The study shows AI is advancing fast. Yet, true artificial general intelligence, or AGI, remains unclear.