AI Fooled Humans and Passed the Turing Test — It Can Now Perfectly Mimic People

Researchers from the University of California, San Diego, have discovered that some cutting-edge neural networks are remarkably adept at masquerading as humans.

In an experiment rooted in the classic Turing Test, models like GPT-4.5 and Llama 3.1–405B showcased impressive prowess during a standard three-way evaluation, while GPT-4o fell short of the mark.

Participants engaged in five-minute conversations with either a human or one of these systems, then attempted to discern who—or what—they’d been chatting with. The exchanges took the form of casual small talk, with a particular focus on the emotional nuances of communication. Participants were free to probe their собеседник with questions about opinions, feelings, and personal experiences.

The results were striking: GPT-4.5 was mistaken for a human more often than an actual person, achieving a convincing 73% success rate. Llama 3.1–405B followed with a solid 56%, while GPT-4o lagged behind at just 21%.

In the Turing Test, if a participant can’t tell whether they’re interacting with a human or a machine, the machine is deemed to have passed. By this measure, GPT-4.5 and Llama 3.1–405B have demonstrated a remarkable ability to emulate human interaction.

The Minds Behind the Study

Conceived by Alan Turing in 1950, the Turing Test gauges a machine’s capacity to mimic human intelligence. When its responses become indistinguishable from those of a person, it’s considered to have crossed the threshold of success.