An A.I. Fooled Humans and Passed the Turing Test. But It's a Red Herring for the Singularity.

Yahoo

10-04-2025

The Turing test has long been an important threshold in evaluating machine intelligence, and OpenAI's latest LLM GPT-4.5 just aced it.
Scientists from the University of California San Diego surmise that current LLMs likely possess the ability to replace humans for short-term conversations, which could cause further job automation and improved 'social engineering attacks,' among other things.
While an impressive engineering feat, this doesn't mean we've achieved artificial general intelligence (AGI). But it does show that humans might be easier to fool than we originally thought.
Even in 1950, at the dawn of the computing age, famous British mathematician and computer scientist Alan Turing knew that machines would one day rival the conversational abilities of humans. To illustrate this idea, Turing developed his eponymous Turing test to gauge whether a machine has become syntactically indistinguishable from its flesh-and-blood creators.
In the ensuing decades, the Turing test has often been touted as an all-important benchmark for the capabilities of advanced computers and AI. And in a recent test, participants mistook GPT-4.5, the latest OpenAI large language model (LLM), for a human 73 percent of the time—far above the 50 percent rate for random chance. A paper discussing the results of this test were uploaded to the preprint server arXiv by scientists at the University of California (UC) San Diego late last month.
'The results constitute the first empirical evidence that any artificial system passes a standard three-party Turing test,' the authors wrote. 'The results have implications for debates about what kind of intelligence is exhibited by LLMs, and the social and economic impacts these systems are likely to have.'
While no doubt impressive, GPT-4.5 had a few tricks up its sleeve to pass itself off as human. First, the authors instructed the LLM to adopt a 'humanlike persona,' which essentially resulted in texts full of internet shorthand and socially-awkward responses. When using this persona, the LLM scored the highest, but without the persona, GPT-4.5 was much less convincing, with an only 36 percent success rate.
These results were conducted in a three-party test, meaning that participants spoke with a human and AI simultaneously and tried to identify which was which. Cameron Jones, a co-author of the study, described this kind of test (which lasts around five minutes) as the 'most widely accepted standard' version of the Turing test in a post on X, formerly Twitter.
While an impressive engineering feat, passing the Turing test is not an indicator that we've officially developed artificial general intelligence (AGI)—the holy grail of the AI world. The Turing test only evaluates one type of intelligence, and some argue that humans possess upwards of nine distinct intelligences (including things like interpersonal, intrapersonal, visual-spatial, and existential). It's for this reason (among others) that some consider the Turing test to be largely obsolete.
However, some people think this milestone represents something more about humans than it does for LLMs. The paper notes, for example, that many participants chose GPT-4.5 based on vibes rather than logic, relying on emotions and feeling rather than asking factual questions or investigating the LLM's reasoning. John Nosta, founder of the think tank NostaLab, wrote in Psychology Today that the Turing test has essentially been 'inverted':
It's no longer a test of machines, it's a test of us. And increasingly, we're failing. Because we no longer evaluate humanity based on cognitive substance. We evaluate it based on how it makes us feel. And that feeling—the 'gut instinct,' the 'vibe'—is now the soft underbelly of our discernment. And LLMs, especially when persona-primed, can exploit it with uncanny accuracy.
Although this test doesn't represent the long-hypothesized moment of singularity when artificial intelligence evolves beyond our own, Jones said on X that it's likely that LLMs can now successfully substitute for people in short conversations, leading to 'automation of jobs, improved social engineering attacks, and more general societal disruption.'
That's why it is important—now more than ever—to regulate the development of AI, or at least approach AI development with immense caution. Unfortunately, the U.S. government currently has no appetite for throttling AI's growing humanlike ambitions.
You Might Also Like
The Do's and Don'ts of Using Painter's Tape
The Best Portable BBQ Grills for Cooking Anywhere
Can a Smart Watch Prolong Your Life?