In an experiment conducted by researchers at the University of California, San Diego, the GPT-4.5 model successfully passed a simplified version of the Turing test in live text-based interaction. Observers identified the model as a "real human" more often than actual humans. The experiment followed a three-way scheme: judges interacted simultaneously with a human and an AI, attempting to identify who was who. This was reported by Ixbt.com reports.
The evaluation was based solely on the content of the correspondence. When assigned a specific "persona," the GPT-4.5 model was identified as human in 73% of cases. For comparison, Meta's LLaMa-3.1-405B model was perceived as human in 56% of cases under the same scenario. Researchers noted that the addition of personalized context made the models' responses significantly more convincing.
The classic Turing test is traditionally used to check an algorithm's ability to mimic human conversation. However, in this study, it served not as a measure of "intelligence," but as an indicator of behavioral credibility in communication. Experts emphasize that the models do not possess consciousness, but rather demonstrate the ability to replicate socially convincing speech patterns in short interactions.
The results show that distinguishing between humans and AI in text chats is becoming statistically difficult. This is significant for customer support, educational platforms, and social networks. The study authors noted the need for clear labeling of AI systems in communication interfaces, stating that transparency has become a social infrastructure issue rather than just a technical one.
Read “Zamin” on Telegram!
