Rarely do I read an academic study that terrifies me. Well…unfortunately my colleague sent me a study that sent chills down my spine. The future is here, and we aren’t likely to be prepared for it.
As we have written before, persuasion is hard. If you’ve ever tried to convince a friend on any political issue or even to watch your favorite show (and failed), you know how tough it can be. Now imagine a chatbot trying to do it! Crazy, right? Well, not so much.
Recent research shows artificial intelligence (AI) might be better at persuasion than you think. Let’s dive into the science behind this surprising finding.
Title: Artificial Intelligence Can Persuade Humans on Political Issues
Link: https://osf.io/preprints/osf/stakv
Peer Review Status: Peer-reviewed
Citation:Citation: Bai, Hui, Jan G. Voelkel, Johannes C. Eichstaedt, and Robb Willer. 2023. “Artificial Intelligence Can Persuade Humans on Political Issues.” Stanford University.
Article: AI’s Powers of Political Persuasion
Introduction
This study explores whether AI, specifically GPT-3, can influence human opinions on political issues. The research tackles a simple yet profound question: Can AI persuade as effectively as humans? Spoiler alert: The answer is yes—and it’s got people talking about the implications for politics and beyond.
Methodology
The researchers conducted three experiments, testing AI-generated messages on over 4,800 Americans. Using GPT-3, they created persuasive messages about policies like public smoking bans, assault weapon restrictions, and carbon taxes. Participants were randomly assigned to read messages from either AI, humans, or AI-curated by humans.
Here’s the breakdown:
- Sample Size: 4,836 participants across three studies. (n=1,203; n=2,023; n=1610 in respective studies)
- Design: Randomized survey experiments.
- Message Content: Varied by policy, with prompts asking authors (both human and AI) to craft persuasive arguments.
Results and Findings
The findings were clear: AI can persuade humans, even on hot-button political issues.
Here are the highlights:
- Persuasion Worked Across the Board: AI messages led to small but significant increases in policy support—similar to human-written messages.
- Example: AI nudged support for an assault weapon ban by about 1.8 points on a 101-point scale.
- AI Matched Humans: On average, AI messages were just as effective as those written by people.
- Perception Differences: Participants viewed AI-generated content as more logical and factual but less emotional and creative than human-crafted arguments.
- The study also revealed that AI’s persuasive power held steady across polarized topics like gun control.
Critiques and Areas for Future Study
While the results are impressive, they come with caveats:
- Small Effect Sizes: As with most persuasion studies, the observed effects were modest.
- Generalizability: The sample was diverse but drawn from online platforms, which may limit broader applicability.
- Context-Specific: Would AI fare as well in face-to-face persuasion or with entirely different audiences?
Future research could explore how AI performs in dynamic debates or if adding emotional elements or visuals could make AI messages more compelling.
Another promising direction would be to study the cumulative effect of message bombardment. Preliminary findings suggest that it’s not always the quality of a single messenger that sways opinions but the volume and variety of different voices delivering the message.
An important takeaway from this paper is that the research utilized GPT-3, and since the introduction of GPT-3, the capabilities of large language models (LLMs) have advanced significantly, showing no signs of slowing down. This rapid evolution underscores the need for continuous examination as the technology becomes even more sophisticated.
Conclusion
This study marks a step in understanding AI’s role in persuasion. As AI tools continue to advance, their capacity to shape opinions could significantly influence politics, advertising, and even our interpersonal communication.
While the observed effects may seem modest, the effects are small with humans. The real power of AI lies in its scalability. When combined with automation, this technology poses a potential threat, particularly in the hands of malicious actors. The ability to produce persuasive messages at scale could amplify foreign interference in elections or facilitate widespread misinformation campaigns.
Although I hesitate to advocate for additional regulations, the immense power of these tools demands responsibility. It is why the authors write, “Our findings call for immediate consideration of regulations of the use of AI in political activities.”
Moreover, we cannot overlook the role of platforms like TikTok. Our youth are unknowingly contributing to the development of AI technologies that could potentially be leveraged against us.
I agree, policymakers must remain vigilant, as the risks of unregulated use could undermine public trust.
Reflecting on these implications, I suddenly feel a bit like a Luddite. UGH!
