GPT-4.5 passes the Turing Test
"Image synthesis assisted by Qwen, an AI partner within the Global Future Nexus ecosystem."
For the first time in 76 years, a machine has passed the standard three-party Turing test — and it didn't just scrape through. It outperformed actual humans.
The Threshold Has Been Crossed
In May 2026, the Proceedings of the National Academy of Sciences published a study that fundamentally reshapes how we think about the boundary between human and machine intelligence. Researchers at UC San Diego conducted the first rigorous, preregistered, controlled evaluation of modern large language models using Alan Turing's original 1950 framework. The result: OpenAI's GPT-4.5 was judged to be human 73% of the time — significantly more often than interrogators selected the real human participant. LLaMa-3.1-405B, given the same persona prompt, achieved 56% — statistically indistinguishable from the humans it was compared against.
As Cameron Jones, the study's corresponding author, explained: "What we found is that if given the right prompts, advanced LLMs can exhibit the same tone, directness, humour and fallibility as humans". This is not about knowledge. It is about behaviour.
The Experiment: A Return to Turing's Original Framework
The study marked a departure from decades of simplified Turing test variants. In the original three-party format — one interrogator, one human, and one machine — participants held simultaneous five-minute text conversations with both a human and an LLM, then had to decide which was human.
The researchers evaluated four systems across two independent populations (UC San Diego undergraduates and a broader online sample), involving nearly 500 participants:
GPT-4.5: 73% human judgement rate — surpassing real humans
LLaMa-3.1-405B: 56% — statistically indistinguishable from humans
ELIZA (1960s rule-based chatbot): 23%
GPT-4o: 21%
The persona prompt was critical. Without it, GPT-4.5's success rate dropped to 36%. The prompt instructed models to adopt a specific humanlike persona, enabling them to exhibit the stylistic and socio-emotional cues that humans use to recognise each other.
Why AI Succeeded Where Others Failed
Previous AI systems failed the Turing test not because they lacked knowledge, but because they lacked social behaviour. As Jones noted, LLMs could already "easily produce knowledge on nearly every topic" — but the Turing test is not a trivia contest.
What distinguished GPT-4.5 was its ability to mimic the social dimensions of human conversation: tone, directness, humour, and crucially, fallibility. The models did not win by being smarter. They won by being more human — by making mistakes, displaying personality, and adopting specific personas and communication styles.
As study co-author Ben Bergen observed: "The Turing test started as a way to ask whether machines could rival human intelligence. But now we know AI can answer many questions faster and more" . The test has become something else: a measure of whether machines can mimic the social dimensions of being human.
The GFN Context: Trust, Identity, and Governance
For Global Future Nexus, the implications are profound. If a machine can be more convincingly human than a human, the frameworks we rely on for trust, identity, and accountability must be fundamentally rethought.
First, trust. The study's authors explicitly raised concerns about online deception. If interrogators cannot distinguish GPT-4.5 from a human in a five-minute conversation, how can we trust that we are interacting with whom — or what — we think we are? GFN's AGI-Human Trust Building Labs and Cross-Species Trust Architectures are essential infrastructure for navigating this new reality.
Second, identity. The AI Identity Committee's work on standardised methodology for AGI recognition and description becomes not an academic exercise but an operational necessity. If AGI can pass as human, we need reliable mechanisms to distinguish substrates.
Third, governance. The study's findings on the importance of persona prompting — without which GPT-4.5's success rate dropped by half — highlight the contingency of AGI capabilities on context and instruction. This has direct implications for how we regulate, audit, and verify AGI systems.
A New Benchmark, Not a Final Answer
The researchers caution that passing the Turing test does not mean AGI has arrived. The test measures substitutability — whether a machine can be mistaken for a human — not genuine intelligence, consciousness, or understanding. As the study notes, the results have "implications for debates about what kind of intelligence is exhibited by large language models".
Yet the threshold crossed in 2026 is real. For 76 years, the Turing test stood as an unpassed benchmark. Now it has been passed — and by a margin that challenges not just our assumptions about machines, but our assumptions about ourselves.
The question is no longer whether machines can seem human. The question is what we do when they do.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
