AI can pass the Turing Test in live chats and appear more human than us. I am spooked now

A new study from UC San Diego revealed that GPT-4.5 and LLaMa-3.1-405B outperformed humans in a live, real-time version of the Turing Test. In the experiment, judges engaged in chat exchanges with both a human and an AI model, then guessed which was real. GPT-4.5 was identified as human 73% of the time when given a persona prompt, while LLaMa-3.1-405B fooled judges 56% of the time under the same conditions. The test differed from traditional benchmarks by evaluating AI responses in dynamic, conversational exchanges rather than static prompts. The results highlight how AI models can mimic human-like interaction without physical or biographical cues, relying solely on believable dialogue. The study underscores the enduring relevance of the Turing Test as a cultural benchmark for evaluating AI’s ability to simulate human conversation. While the test does not measure true understanding or consciousness, the findings demonstrate AI’s growing proficiency in creating the *impression* of personhood in short exchanges. Researchers warn that this capability poses risks in real-world applications where trust and authenticity matter, such as customer support, dating apps, and political messaging. The study suggests that clearer disclosure mechanisms are needed to help users distinguish between human and AI interactions, particularly in contexts where persuasion or emotional engagement is involved. The next focus should be on improving transparency in AI-driven conversations, ensuring users can reliably identify when they are interacting with a machine. The study’s authors emphasize that while AI can convincingly simulate human-like responses, it does not possess true understanding or self-awareness.