GPT-4.5 passes the Turing test when playing a persona

For decades, passing the Turing test was treated as the holy grail of artificial intelligence. It turns out that finally beating it didn't require complex mathematical reasoning or conscious thought. The AI just had to pretend to be an internet-savvy kid.

In a new preprint study awaiting peer review, researchers at UC San Diego’s Language and Cognition Lab ran standard, pre-registered three-party Turing tests with nearly 300 participants. Interrogators chatted with a human and an AI simultaneously for five minutes, then had to declare which was the real person.

The results were striking: OpenAI's GPT-4.5 was judged to be human 73% of the time when given a specific persona. That didn't just beat random chance at 50%; it meant human judges picked the AI as the human significantly more often than actual human participants.

The magic of persona prompting

The researchers put four systems through the imitation game: OpenAI's GPT-4.5 and GPT-4o, Meta’s open-weight LLaMA 3.1-405B, and ELIZA, the rudimentary chatbot created roughly eight decades ago.

The secret weapon wasn't model scale alone; it was prompt framing.

When given a generic prompt telling the AI to convince the judge it was human, GPT-4.5 won only 36% of the trials. GPT-4o fared even worse at 21%, marginally trailing ELIZA’s 23% win rate. Without a specific personality, the models sounded sterile and machine-like.

Everything changed once the models received a persona prompt directing them to act like a young person steeped in internet and casual culture. GPT-4.5 leaped to a decisive 73% win rate, while Meta’s LLaMA 3.1-405B cleared the bar at 56%, leaving judges no better than chance at telling it apart from real humans.

Small talk fooled the judges

If you thought Turing test interrogators would grill the machines on quantum physics or logic puzzles, think again. Over 60% of the conversation logs focused on mundane daily activities, personal routines, and emotional vibes.

Human interrogators searched for authenticity through casual banter. Because today's frontier LLMs are trained on vast reserves of human chatter, adopting colloquial quirks and casual humor proved trivial for them.

Lead author Cameron Jones noted that this marks the first empirical proof of artificial systems winning a standard three-party test. But as Google engineer François Chollet previously told Nature, the imitation game was originally conceived as a philosophical thought experiment rather than an operational evaluation benchmark.

Why it matters

First, this study effectively retires the Turing test as a useful benchmark for frontier AI intelligence. LLMs have demonstrated that convincing mimicry can pass an evaluation that many once assumed required true human thought. Passing the test tells us a lot about conversational fluency and prompt engineering, but very little about underlying machine cognition.

Second, the practical consequences for everyday digital interaction are immediate and messy. If people cannot tell a machine from a human across five minutes of casual text exchange, the barrier for convincing impersonation is completely gone.

We think this highlights an impending security headache. While automated customer support and personalized agents will become much smoother, the flip side is that social engineering attacks, synthetic scams, and astroturfing will become vastly harder to detect. Distinguishing people from bots on text platforms just shifted from an academic curiosity to an urgent verification crisis.