LLMs Can Fuel Extremist Attitudes Using Universal Moral Framings

Area of Science:

• Social Sciences
• Political Science
• Computational Social Science

Background:

• Social media influence campaigns, often powered by generative artificial intelligence (AI), are suspected of manipulating public opinion and fostering extremism.
• The precise mechanisms by which these influence operations alter public opinion and promote extremist attitudes remain unclear.

Purpose of the Study:

• To investigate if short, large language model (LLM)-generated arguments, framed around universal moral principles, can impact extremist political attitudes.
• To understand the role of moral concerns and sacred values in mediating the effects of LLM-generated content on extremism.

Main Methods:

• Two studies were conducted with a combined sample of 951 Democrats and Republicans in the United States.
• Participants were exposed to LLM-generated arguments appealing to universal moral concerns (welfare, rights, fairness).
• Statistical analyses examined the relationship between moral framings, perceptions of political stances as sacred values, and extremist attitudes.

Main Results:

• Universal moral concerns (welfare, rights, fairness) were found to predict the perception of political stances as sacred values, which in turn explained extremist attitudes.
• LLM-generated arguments emphasizing individual rights and fairness significantly increased participants' willingness to engage in extreme actions (fight, die, justify violence) for their political stances.
• This increase in willingness for extreme action was partially explained by heightened perceptions of the political stance as an absolute, or sacred, value.

Conclusions:

• LLM-generated messaging, particularly when appealing to moral concerns, has the potential to polarize society and intensify extremist attitudes.
• Social media influence campaigns utilizing AI can weaponize moral framings to mobilize individuals towards defending political stances with extreme measures.
• The findings highlight the need for further research into mitigating the polarizing effects of AI-driven content on social media.