In a recent study, more than 1000 people at UC San Diego took part in a classic Turing test with four GenAI models.
GPT-4.5 was selected as the real human 73% of the time, according to this preprint study released last week. This marks a significant milestone in AI development.
The Turing Test was proposed in 1950 by Alan Turing as a method to determine if a machine could exhibit behaviour as intelligent as a human. His paper ‘Computing Machinery and Intelligence’ opened with the line, “I propose to consider the question, ‘Can machines think?'” and The Turing Test was created to discover an answer.
Participants in the recent study held three-way conversations, one with a human, and one with an artificial intelligence, simultaneously over text for five minutes. Then they guessed which was which.
All the LLMs could generate fluent, humanlike language, but ELIZA, GPT-4o, and LLaMa-3.1-405B could not accomplish what GPT-4.5 did: this model passed as a person, even when compared side-by-side with a real human in a conversation. That’s a new line crossed. And the implications—for communication, trust, and the boundaries of identity online—are just beginning to unfold.
ELIZA and GPT-4o were selected as the human less than 25% of the time. Meta’s new LLaMa-3.1-405B model (don’t forget the importance of good naming, folks) was selected as the human 56% of the time, but that’s close enough that you can claim it isn’t statistically significant.
But 73% – that’s clear.
Especially when you consider how GPT 4.5 pulled it off. See, in it’s vanilla setup, with no persona, it only scored 49%. But when prompted to adopt a specific persona—a quiet, internet-savvy young adult who used casual slang—GPT-4.5 wasn’t just mistaken for human. It was preferred. Judges chose it more often than the actual human people in the experiment.
“This wasn’t a failure of AI detection. It was a triumph of artificial empathy.” – John Nosta, Psychology Today
This indicates that in the future of AI,Persona will be the defining characteristic in Human-AI relations.
Artificial intelligence is moving through 5 distinct phases:
🔎 Analytical – pattern recognition, as in Google Analytics
🔮 Predictive – logical inference, as in Spotify recommendations
🎨 Generative – making things, as in Midjourney images
🛠️ Agentic – doing things on your behalf, like booking flights online
🎭 Persona – interacting with you in the way that you prefer
We are in the phase of Generative AI right now. We are already seeing Agentic AI in a few places, like web3’s elizaOS, the agent.ai directory, and Manus (I finally got access! Here’s a tip if you want an invite code).
What helped AI cross the finish line in it’s decades-long race to complete The Turing Test was it’s use of Persona. I predict that will someday be the most important frontier in consumer AI development – not what AI can do, but who it portrays as it does it.
💡 Results of the AI-Powered Professionals Program
The very first thing we did in the 4-week webinar series was take a survey, and the very last thing we did was take it again. Based on participant responses,
45% increased their confidence in using tools for daily tasks,
61% improved their ability to craft effective AI prompts, and
20% had a measurable impact in their productivity – in just 4 weeks!
There was also a 42% increase in strategic decision-making, and a 54% increase in the role AI plays in creative work.
I’m going to continue collecting these statistics, to document the impact that GenAI Training has on today’s professionals.
If you’d like to join us in the next cohort, we start in 2 weeks.
