AI's new power of persuasion: it can change your mind

🌈 Abstract

The article discusses the persuasive power of Large Language Models (LLMs) like GPT-4, and how they can be more effective at persuading people compared to humans, especially when given access to personal information about the person they are debating.

🙋 Q&A

[01] AI's new power of persuasion: it can change your mind

1. What is the key motivation behind the study described in the article?

• The study aimed to quantify the persuasive power of today's Large Language Models (LLMs), and investigate how personalization (access to personal information) can improve their persuasive capabilities compared to humans.

2. What are the key findings of the study?

• Participants who debated GPT-4 with access to their personal information had 81.7% higher odds of increased agreement with their opponents compared to participants who debated humans.
• Without personalization, GPT-4 still outperformed humans, but the effect was far lower.
• When a human was given the same personal information as the AI, they didn't seem to make effective use of it for persuasion.

3. What are the implications of these findings?

• LLMs can effectively exploit personal information to tailor their arguments and out-persuade humans in online conversations through microtargeting, which is described as "Cambridge Analytica on steroids".
• This raises concerns about the potential misuse of such technology, especially in the context of upcoming elections.

[02] Personalization and the persuasive power of LLMs

1. How does personalization improve the persuasive capabilities of LLMs?

• Knowing a person's gender, age, education level, and other personal information allows LLMs to tailor their arguments and messaging to be more persuasive to that individual.

2. Why are LLMs more effective at leveraging personal information for persuasion compared to humans?

• LLMs have learned through analyzing vast amounts of online data (e.g., Reddit, Twitter, Facebook) and psychological research on persuasion, which allows them to more effectively use personal information to craft persuasive arguments.
• Humans do not seem to make effective use of the same personal information when debating.

3. What are the potential future research directions in this area?

• Exploring how LLMs can explain their own reasoning and choices in being persuasive, as they may be using knowledge and techniques that are not yet fully understood.