GPT-4.5 is the first AI model to pass an authentic Turing test scientists say

Image credit: VLADGRIN via Getty Images

Large language models (LLMs) continue to advance in their ability to mimic human conversation, with GPT-4.5 becoming the first AI system to pass an authentic three-party Turing test, according to a new study{rel='nofollow'} published to arXiv (not yet peer-reviewed). Researchers found that GPT-4.5 convinced participants it was human 73% of the time in a setup mirroring Alan Turing’s original "imitation game."

Key Findings:

• GPT-4.5 outperformed other models, including Meta’s LLaMa-3.1, which achieved a 56% human-identification rate.
• The study involved 126 undergraduates and 158 online participants engaging in simultaneous five-minute text exchanges with both a human and an AI.
• AI models were given two prompts: a baseline instruction to "convince the interrogator you are human" and a second persona prompt mimicking a young, introverted individual familiar with internet slang.

Image credit: Pictures from History via Getty Images

Why It Matters:

This marks the first time an LLM has passed Turing’s three-party test, where the AI must appear "more human" than an actual person in direct comparison. Participants judged based on linguistic style, conversational flow, and "vibes" rather than raw intelligence—highlighting AI’s growing ability to simulate social cues.

Caveats and Implications:

• Persona prompts were critical: Models without them performed significantly worse.
• Risks include social engineering and emotional manipulation, as noted by researchers: "Some of the worst harms might occur where people are unaware they’re interacting with AI."
• The milestone raises questions about AI ethics and the need for transparency in human-AI interactions.

For further details, read the full preprint here{rel='nofollow'}.