GPT-4.5 and LLaMa-3.1 Excel in Turing Test Performance
LLMs and the Turing Test: GPT-4.5 and LLaMa-3.1 Performance
Recent studies have evaluated the performance of Large Language Models (LLMs) in the Turing Test, specifically focusing on GPT-4.5 and LLaMa-3.1. The results provide significant insights into their ability to mimic human-like conversation.
Key Findings
- GPT-4.5 with Persona Prompt: Achieved a win rate of 73%, meaning it was judged to be human 73% of the time. This performance was significantly higher than chance and even outperformed real human participants in some cases.
- LLaMa-3.1 with Persona Prompt: Achieved a win rate of 56%, which was not significantly different from the human baseline, indicating that interrogators could not reliably distinguish it from a human.
- Baseline Models (ELIZA and GPT-4o): Performed significantly below chance, with win rates of 23% and 21% respectively, confirming that interrogators could easily identify these models as non-human.
Implications
The results suggest that advanced LLMs like GPT-4.5, when prompted to adopt a humanlike persona, can pass the Turing Test. This has profound implications for debates on artificial intelligence, particularly regarding the social and economic impacts of systems that can convincingly mimic human behavior.
Study Details
The study involved randomized, controlled, and pre-registered Turing Tests with independent populations. Participants engaged in 5-minute conversations with both a human and an AI system before judging which was human. The findings highlight the importance of persona-based prompts in enhancing the human-like quality of AI responses.
For more detailed information, you can access the full study here.
Invite Gmail Contacts
