Topics such as the war in Ukraine and current attacks on Israel are just a few examples where insults, hostility and hate comments can be found in social media and comment columns. According to a survey conducted in the summer on behalf of the Media Authority of North Rhine-Westphalia, 76 percent of all respondents have already experienced hate speech online, and just under 40 percent have been affected themselves. These results are in line with surveys from the USA.
In terms of improving the sometimes toxic discourse on the Internet, researchers from the U.S.’s Brigham Young University and Duke University have a suggestion – the use of artificial intelligence (AI) to make conversation on the Internet more polite and friendly. Their findings were recently published in the journal PNAS.
The study focused on a specially designed online platform that was tested with 1,500 people. On this platform, two people at a time with different views on an emotionally and politically charged topic, in this case gun ownership in the U.S., came together anonymously. They were encouraged to discuss the topic, but were moderated by an AI chatbot in the background.
The AI assistant, based on the GPT-3 language model, was able to provide customized suggestions for rephrasing immediately after comments entered to make responses more polite or friendly, but without distorting their meaning or thesis. Participants had the option of accepting, rejecting, or revising each AI suggestion themselves. At the end of the test, they were asked to rate the quality of the conversation.
According to David Wingate, computer science professor and a contributor to the study, the more often the rephrasings were used, the better participants felt they understood. The suggestions were based on three techniques from standard discourse and communication theory: repetition, validation and politeness. During the test, a total of 2,742 rephrasings were performed by the AI and two-thirds of the participants accepted the suggestions. Forty percent of the rephrasings were pleasantries, while 30 percent each were based on repetition and validation.
According to the study, the results suggest that this simple intervention can improve the quality of conversation and strengthen democratic reciprocity. The researchers see potential to scale the system, such as applying it to the comments column of an online publication or to platforms like Facebook. This could relieve the burden on human moderators.
However, not everything is that simple. The team that developed the AI for Better Phrasing emphasizes that their AI is capable of identifying personal viewpoints without distorting opinions through rephrasing. But caution is warranted. Large language models like GPT are not free of bias, and depending on the implementation, they may favor certain viewpoints.