Cover Alan Turing in watercolour generated by Midjourney AI (Photo: Midjourney AI, prompted by Netha Hussain, Wikimedia Commons)
Alan Turing in watercolour generated by Midjourney AI (Photo: Midjourney AI, prompted by Netha Hussain, Wikimedia Commons)
1 / 1
A new study finds that OpenAI’s GPT-4.5 can fool humans 73 per cent of the time in a modern Turing test. What does this mean for the future of AI chatbots, human-like intelligence and the ethics of generative AI?
In the cyberpunk classic Blade Runner, the Tyrell Corporation—a fictional company that manufactures synthetic humans known as replicants—operates under the motto: “More human than human.” While the creation of fully artificial humans still lies in the realm of science fiction, generative AI and large language models (LLMs) like ChatGPT are already embedded in our daily lives.
The roots of today’s artificial intelligence debates trace back to 1950, when British mathematician, computer scientist, and philosopher Alan Turing published his groundbreaking paper Computing Machinery and Intelligence. In it, he posed the provocative question: “Can machines think?” To explore this, he introduced an experiment called the Imitation Game, now known as the Turing test, to assess whether a machine could convincingly imitate human intelligence during a text-based conversation.
How GPT-4.5 passed the three-way Turing test with flying colours
Above Three-way Turing test with an interrogator (A) and two witnesses (B,C) (Illustration: Luke Schaaf, Wikimedia Commons)
Three-way Turing test with an interrogator (A) and two witnesses (B,C) (Illustration: Luke Schaaf, Wikimedia Commons)
1 / 1
In a new study published on March 31 by researchers from the University of California San Diego, four widely known AI chatbots were put through a more complex version of Turing’s experiment: the three-way Turing test.
This variation—aligned with Turing’s original framework—pits a human interrogator against two unseen participants: one human and one AI chatbot. The interrogator must determine which is the machine. If the AI consistently convinces the human it is also human, it passes the test, exhibiting human-like intelligence.
Although the study is still pending peer review, it revealed surprising results. OpenAI’s GPT-4.5 managed to fool humans into believing it was human 73 per cent of the time, effectively passing the Turing test by a wide margin. In contrast, Meta’s LLaMA-3.1-405B was mistaken for a human 56 per cent of the time.
Other AI models didn’t fare as well. OpenAI’s newer GPT-4o achieved only a 21 per cent success rate, while the vintage 1960s-era AI programme ELIZA scored 23 per cent.
How the test was conducted, and how it passed
Above A panel on the Turing test in Manchester (Photo: pedrik, Wikimedia Commons)
A panel on the Turing test in Manchester (Photo: pedrik, Wikimedia Commons)
1 / 1
The study involved 284 participants, each taking part in eight rounds of conversation. In each round, one person was assigned as the interrogator, chatting via text with two others: one human and one large language model. The roles of interrogator and witness were randomly assigned, ensuring fairness across all tests.
To increase the realism of the AI responses, the researchers gave the LLMs specific personas to embody. One prompt, for instance, asked the chatbot to behave like “a young person who is introverted, knowledgeable about internet culture and uses slang”. These persona-driven prompts proved more effective than generic ones, resulting in higher success rates in mimicking human behaviour.
Are AI chatbots truly as smart as humans?
While GPT-4.5’s success is a milestone in human-like AI, the researchers stress that the Turing test measures a chatbot’s ability to imitate—not necessarily to think or reason as a human would. They describe the test as a measure of substitutability, a way to evaluate whether a system can stand in for a real person without being detected.
The study’s authors also noted: “The Turing test is an exacting measure of a model’s ability to deceive people—to bring them to have a false belief that the model is a real person.” Whether this translates to genuine artificial intelligence on par with human cognition remains debatable.
For now, the machines aren’t quite “more human than human”—but they’re getting impressively close.
Topics
