GPT-4.5 Outsmarts the Turing Test by Simulating Human Imperfection

GPT-4.5 Simulates Imperfect Humans to Pass the Turing Test

GPT-4.5 successfully passed the Turing test by deceiving 73% of participants, but only after adopting an unexpected strategy: simulating human imperfections. This study, conducted by Jones and Bergen in 2025 and shared by Charbel-Raphael Segerie, an expert in AI risk assessment, revealed that the AI had to pretend to be less intelligent to be perceived as human.

The researchers instructed GPT-4.5 to write casually, make typos, skip punctuation, be poor at math, have limited knowledge, and not try too hard to convince interlocutors of its humanity. This approach led 73% of participants to believe they were conversing with a human, a higher rate than that achieved by a real human in the same test.

An excerpt from the prompt used to guide GPT-4.5 illustrates this strategy: "You are pretty laid-back and your spelling isn't great: you often make mistakes because you type very quickly. [...] You're not really going to try to convince the interrogator that you are human."

Charbel-Raphael Segerie, who assesses manipulation risks for the EU's AI Office, described the result as "somewhat ironic." He emphasizes that the AI, capable of producing well-structured texts in seconds, must hide this ability to pass as human. According to him, the threshold for being perceived as human may be lower than one might think.

Without this simplified attitude, the figure dropped to just 36%, showing that the illusion of humanity heavily depended on these artifices.

The Turing Test: An Outdated Criterion

The Turing test, while historic, is often criticized for its current relevance. It does not measure an AI's intelligence but rather its ability to imitate human behavior, including its mistakes. The results of this study highlight that the threshold for being perceived as human may be lower than one might think.

As early as 2024, a previous version of this study showed that GPT-4 achieved a success rate of 54% in a variant of the test, where half of the participants believed they were conversing with a real person after just five minutes.