Dialogues with large language models reduce conspiracy beliefs even when the AI is perceived as human

Area of Science:

• Psychology
• Artificial Intelligence
• Communication Studies

Background:

• Conspiracy beliefs are often resistant to correction.
• Recent studies indicate large language models (LLMs) can reduce these beliefs through personalized dialogues.
• The effectiveness of AI persuasion in correcting beliefs is not fully understood, particularly regarding the messenger's identity.

Purpose of the Study:

• To investigate whether the effectiveness of AI-driven belief correction relies on the AI messenger or the persuasive content.
• To compare the impact of an AI versus a human messenger delivering fact-based counterarguments.
• To examine the role of conversational tone in AI persuasion.

Main Methods:

• A preregistered experiment with 955 participants was conducted.
• Participants reported either a conspiracy belief or an epistemically unwarranted belief.
• Participants interacted with an LLM that argued against their belief, with conditions varying the LLM's characterization (AI vs. human expert) and conversational tone.

Main Results:

• LLM-driven conversations significantly reduced participants' confidence in both conspiracy beliefs and epistemically unwarranted beliefs.
• No significant differences in belief reduction were observed across the different messenger conditions (AI vs. human) or conversational tones.
• The effectiveness of the LLM's persuasion was independent of whether it was perceived as an AI or a human expert.

Conclusions:

• AI persuasion in correcting unwarranted beliefs is effective due to the compelling nature of the generated messages, not the AI identity of the messenger.
• Fact-based arguments delivered through AI are as effective as those delivered by human experts.
• Future research should focus on optimizing message content for belief correction, irrespective of the messenger's nature.