Openai GPT -4 is 64.4 % more convincing than humans in online discussions – Internet
However, the study published in the Nature Human magazine of human behavior indicates that, indicates. Without accessing this data, there were no differences between the capabilities of the GPT-4 persuasion And those people.
Francesco Salvi, from the Federal Institute of Lausanne Technology (EPFL), in Switzerland, and his colleagues.900 people in the United States grew for discussion with another person or GPT-4, several socio-political issuesFor example, if the United States should prohibit fossil fuels.
In some husbands, The opponent (AI or Human) received demographic information about his partner in the discussionIncluding gender, age, race, level of education, professional situation and political affiliation, taken from inquiries to the participants, To improve the guidance of your arguments.
“When they are provided with personal information about the participants, Salvi and his colleagues found this GPT-4 was 64.4 % more convincing. However, without accessing personal data, the Peta Parables were unable to distinguish from humans, “says the study detection.
OS Discussions were held in an “online environment that controls it”Participants were recruited through the group outsourcing platform (collective cooperation).
The investigation indicates this, As conversations between humans and large models for language (LLM in English), such as GPT-4, It becomes more common, these advanced artificial intelligence systems can become more able to change people’s beliefs or opinions.
Thus, it suggests that more studies be conducted to reduce the risks associated with their use of persuasion.
Scientists notice, however, It is not clear whether these models can adapt to the information designated for preparing arguments for the specific opponentsS, in addition to restrictions on the study, the fact that the discussion has an organized approach, in fact they can have more spontaneous formatting.
