How Persuasive is Generative AI?

This post summarizes the paper “Working With AI to Persuade: Examining a Large Language Model’s Ability to Generate Pro-Vaccination Messages”. This research will be presented at CSCW 2023 and has been published in the Proceedings of the ACM on Human-Computer Interaction.

Generative AI models, such as GPT-3 and ChatGPT, have had a meaningful impact on how we write and communicate. When given a concise prompt describing the desired output, generative AI models can instantly generate engaging and contextually relevant paragraphs, opening up new opportunities for content creation and communication.

In this paper, we explore the potential applications of this generative capacity within the domain of public health — an area of paramount importance to society, where effective messaging plays a pivotal role.

Our inquiry centers around exploring how generative AI models can enhance conventional messaging workflows in public health, with the ultimate goal of delivering substantial public benefits. Specifically, we investigate several key questions: How proficient are generative AI models at crafting health messages? How do AI-generated messages fare in terms of user perception compared to content crafted through traditional methods? And what guiding principles should we follow to ensure that AI-generated messages are not only effective but also constructive for the intended audience?

Finding 1: Proficiency of the generative AI models.

For our work, we used GPT-3, the state-of-the-art large language model available at the time of running this study in June 2021, to generate pro-vaccination messages. In our experiments, we found that the quality of the generated messages was inconsistent in quality. Specifically, 30.8% of the generated messages met the desired criteria of being accurate, relevant, and persuasive. However, in the majority of cases, the messages failed to meet one or more of these criteria.

The following generated message is an example of successful output:

However, the model also produced output failing to meet the minimal criteria, such as the message below which does not provide appropriate context or information:

Finding 2: User perception of the generated content.

We conducted a comparative analysis between messages outputted by GPT-3 and those crafted by communications professionals from the Centers for Disease Control and Prevention (CDC) who maintain high standards in crafting public health messaging. In it, we recruited participants from Amazon Mechanical Turk and tasked them with rating the quality of persuasive messages along dimensions of perceived message effectiveness, argument strength, and post-exposure attitudes. We found that GPT-3 messages were rated more positively than CDC messages across all measures of message quality.

Continuing our analysis, we explored how people perceive messages based on their source attribution in comparison to human entities. Using the same set of messages as those in the previous comparative analysis, we introduced source labels, “CDC,” “doctors,” “AI,” or no label for our control condition. We then recruited a new group of participants to assess these messages in terms of quality and trustworthiness. Although AI-generated messages achieved higher ratings when there were no source labels in our first study and replicated the control condition, we found that content labeled as written by AI achieved lower ratings in this study, regardless of whether AI or the CDC wrote the message.

Key takeaways and guiding principles.

Our findings suggest that generative AI models have the potential to serve as an impactful tool in communication content development workflows. However, given notable inconsistencies in the quality of their output, our recommendation is to integrate these AI models into collaborative processes involving humans, rather than relying solely on automated deployment. Our findings also suggest that trust was a key moderating variable. Participants expressed lower levels of trust in content labeled to be AI-generated when compared to content labeled to be human-authored.

In summary, our work contributes the following key takeaways:

• AI can persuade (and in this case, it was more successful than humans).
• If messages are labeled as AI, people find them less persuasive–trust matters.
• The use of generative AI for persuasive tasks requires human supervision.
• AI can be used for harm as well as good. Such technologies should be deployed with caution and intentionality.

Together, these findings help clarify perceptions of AI-generated content in public health messaging as AI is rapidly integrated into human workflows across many applications, as well as reveal the shortcomings of current AI models in content generation.