OpenAI’s GPT-4.5 passed the Turing Test with 73% accuracy

In a new study awaiting peer review, OpenAI’s GPT-4.5 outperformed its older version and other chatbots by passing the Turing test—a classic benchmark for human-like intelligence.

Researchers at UC San Diego’s Language and Cognition Lab reported that in the test where participants chat with an AI bot and a person simultaneously and determine which interaction is with a human, 73% of the time the GPT 4.5 model passed as a person.

The result is significantly higher than the 50% random chance.

Other models tested included GPT-4o (which flopped at just 21%) and Meta’s LLaMa 3.1, which also performed on par with ChatGPT. Even the vintage chatbot ELIZA (developed in the 1960s) eked out a 23% score.

However, the chatbot could pass off as human only when given the persona prompt. Without it, GPT-4.5 only fooled 36% of people.