GPT-4.5 coped with three-sided Turing test

Researchers conducted a trilateral turing test for four AI systems-Eliza, GPT-4O, LLAMA-3.1-405B and GPT-4.5. The latter scored the highest score.

In March 31, the work of Cameron Jones and Benjamin Bergen from the Department of Cognitive Sciences of the University of California at San Diego shared the results of the experiment.

They applied the original trilateral version of the test-the participants were conducted by five-minute conversations at the same time as another interlocutor and one of the AI ​​systems, after which they determined which of the interlocutors was considered a person. This option is more difficult compared to the test, where people communicate only with the machine.

In 73% of cases, the subjects considered the GPT-4.5 person. Other AI gained a lesser result:

• Llama-3.1-56%;
• Eliza – 23%;
• GPT-4O-21%.

The Tering test is a conceptual test proposed by the British mathematician Alan Turing in 1950 to determine the ability of a computer to demonstrate intellectual behavior, indistinguishable from the human one.

The essence of the test:

1. A person leads a text correspondence with two interlocutors: another person and artificial intelligence.
2. If the subject cannot with confidence to determine which of them the car is believed that the computer has passed the test.

Test Turing has been repeatedly carried out among popular AI models. So, in June 2024, people were unable to distinguish ChatGPT from a human interlocutor in 54% of cases. Eliza then scored 22%, GPT-3.5-50%, a person-67%.

In 2023, in a similar study from Jones, GPT-4 scored 41%, GPT-3.5-14%, Eliza-27%. People then received 63%.

Recall that in February 2025, Openai released a new version of the GPT-4.5 chatbot with advanced “emotional intelligence”.

Be in the know! Subscribe to Telegram.