As a result of extensive research conducted by scientists at the University of California, it has been determined that artificial intelligence (AI) has the ability to deceive people. Specifically, the GPT-4.5 model was able to convince 73% of participants in the Turing test that it was a real human.
How was the research conducted?
The research was carried out in San Diego. It involved four artificial intelligence systems (LLaMa-3.1−405B, GPT-4o, GPT-4.5, and ELIZA) and 284 participants. After a five-minute conversation, participants were tasked with identifying which of the two interlocutors was a bot.
The test results were as follows:
- GPT-4.5 – convinced 73% of participants that it was a real human.
- LLaMA-3.1 – succeeded in 56% of cases.
- GPT-4o – was accepted as human by people in 21% of cases.
- ELIZA – managed to deceive 23% of participants.
The secrets of GPT-4.5's success
According to scientists, the GPT-4.5 model achieved the highest results through personalization. When AI presented itself as a young introvert aware of internet culture and jargon, people were more likely to trust it. If this adaptation process was not carried out, its success rate dropped to just 36%.
Research conclusions and potential risks
Experts emphasize that such results could lead to certain social risks. Artificial intelligence already has the capability to engage in written communication with people, which could be widely used in customer support services, social networks, and online communication platforms.
Additionally, the emergence of romantic chatbots, fake friends, and automated trolling systems in real life is also being observed. However, there are also positive opportunities — new avenues are opening up for automating jobs that require communication.
Passing the Turing test — is it a sign of "intelligence"?
So far, no AI system has been able to consistently pass the classic three-party Turing test — in this test, AI must directly argue with a human. However, GPT-4.5 became the first artificial intelligence to achieve this. At the same time, scientists emphasize that such success does not mean that AI has remained "intelligent." Successfully passing the Turing test demonstrates not the cognitive abilities of artificial intelligence but its success in imitating humans.
According to the research results, the impact of AI on human life may increase in the future. This indicates the need to develop new rules and laws regarding its application and regulation.
