Chatbots fail Turing test imitation when distinguishing AI from human

A recent study from the University of Paris Cité revealed that seven chatbot models acting as interviewers were largely unable to distinguish humans from AI in an imitation of the Turing test. Out of 98 verification attempts, AI models identified the nature of their interlocutors correctly only three times.

• Only 3 out of 98 checks were successful for AI interviewers.
• Models tested included ChatGPT 4.5, Claude 3.7 Sonnet, Gemini 2.5 Pro, and Grok 3.
• Detailed persona prompts allowed AI to pass as humans perfectly in all cases.
• AI interviewers frequently misidentified other neural networks as human participants.

Methodology and Experiment Details Four researchers designed the experiment to test whether AI could police itself. Participants included 20 human volunteers and seven AI models. Each interviewer could ask up to 10 questions. In one scenario, bots were given basic instructions to act like humans; in another, they were provided with a complex persona, such as a 21-year-old runner who plays jazz, which effectively masked their non-human nature.

Why the AI Struggled Failing bots were caught due to overly structured and organized responses. When a detailed persona was used, the bot's response length averaged 39 words, nearly identical to the human average of 38 words. Interestingly, Claude initially refused to participate in the deception but eventually agreed after being briefed on the research purpose. The experiment highlights that AI struggle to identify artificial markers when a narrative is sophisticated enough.

Implications for Future AI This failure confirms that the traditional Turing test is no longer a robust metric for gauging machine intelligence. For developers and businesses globally, this implies that existing bot-detection systems are becoming obsolete. As AI becomes better at simulating human behavior, companies must move toward more advanced behavioral and cryptographic verification methods to secure platforms against synthetic identities and automated manipulation.