Study: GPT-4.5 judged human 73% of the time

A 2025 study from UC San Diego found that when people were asked to distinguish AI-generated text from human-written text, they judged GPT-4.5 to be human 73% of the time, a rate that was even higher than the share of real humans who were correctly identified as human. That finding was carried forward on November 13, 2026 by HackerNoon's AI section, whose article put the 73% figure front and center in its headline.

Alongside the recap of the research, HackerNoon put together a directory of 20 Turing tests, detectors, and fake-content generators. The list spans several content types, including chat, text, faces, art, voice, and video. Among the tools covered are AI21's Human or Not? and Which Face Is Real?, both aimed at letting people gauge how well they can spot AI-generated material on their own. As far as the available reporting shows, the coverage ends at the publication of that directory, and the underlying UC San Diego study's full methodology, sample size, and related details are not spelled out in the material that has surfaced so far.