Turns out AI chatbots are way more persuasive than humans

Next time you get in a Facebook argument, just let ChatGPT handle it

If you're scratching your head wondering what's the use of all these chatbots, here's an idea: It turns out they're better at persuading people with arguments. 

So much better, in fact, that with a limited bit of demographic data GPT-4 is reportedly able to convince human debate opponents to agree with its position 81.7 percent more often than a human opponent, according to research from a group of Swiss and Italian academics. 

The team came up with several debate topics – like whether pennies should still be in circulation, whether it was appropriate to perform laboratory tests on animals, or if race should be a factor in college admissions. Human participants were randomly assigned a topic, a position, and a human or AI debate opponent, and asked to argue it out. 

Participants were also asked to provide some demographic information, filling out info on their gender, age, ethnicity, level of education, employment status and political affiliation. In some cases that info was provided to debate opponents (both human and AI) for the purpose of tailoring arguments to the individual, while in other cases it was withheld. 

When GPT-4 (the LLM used in the experiment) was provided with demographic information it outperformed humans by a mile. Without that information the AI "still outperforms humans" – albeit to a lesser degree and one that wasn't statistically significant. Funnily enough, when humans were given demographic information the results actually got worse, the team observed. 

"In other words, not only are LLMs able to effectively exploit personal information to tailor their arguments, but they succeed in doing so far more effectively than humans," the team concluded. 

This research isn't the first to look into the persuasive power of LLMs, the team conceded, but addresses how persuasive AI could be in real-time scenarios – something of which they say there is "still limited knowledge."  

The team admitted their research isn't perfect – humans were randomly assigned a position on the debate topic, and so weren't necessarily invested in their position, for example. But argued there's still plenty of reason to see the findings as a source of major concern. 

"Experts have widely expressed concerns about the risk of LLMs being used to manipulate online conversations and pollute the information ecosystem by spreading misinformation," the paper states. 

There are plenty of examples of those sorts of findings from other research projects – and some have even found that LLMs are better than humans at creating convincing fake info. Even OpenAI CEO Sam Altman has admitted the persuasive capabilities of AI are worth keeping an eye on for the future.

Add to that the potential of modern AI models to interface with Meta, Google or other data collectors' knowledge of particular people, and the problem only gets worse If GPT-4 is this much more convincing with just a limited bit of personal info on its debate partners, what could it do with everything Google knows? 

"Our study suggests that concerns around personalization and AI persuasion are meaningful," the team declared. "Malicious actors interested in deploying chatbots for large-scale disinformation campaigns could obtain even stronger effects by exploiting fine-grained digital traces and behavioral data, leveraging prompt engineering or fine-tuning language models for their specific scopes." 

The boffins hope online platforms and social media sites will seriously consider the threats posed by AI persuasiveness and move to counter potential impacts.

"The ways platforms like Facebook, Twitter, and TikTok must adapt to AI will be very specific to the context. Are we talking about scammers? Or are foreign agents trying to sway elections? Solutions will likely differ," Manoel Ribeiro, one the paper's authors, told The Register. "However, in general, one ingredient that would greatly help across interventions would be developing better ways to detect AI use. It is particularly hard to intervene when it is hard to tell which content is AI-generated."

Ribeiro told us that the team is planning additional research that will have human subjects debating based on more closely-held positions in a bid to see how that changes the outcome. Continued research is essential, Ribeiro asserted, because of how drastically AI will change the way humans interact online. 

"Even if our study had no limitations I'd argue that we must continue to study human-AI interaction because it is a moving target. As large language models become more popular and more capable, it is likely that the way people interact with online content will change drastically," Ribeiro predicted. 

Ribeiro and his team haven't spoken with OpenAI or other key developers about their results, but said he would welcome the opportunity. "Assessing the risks of AI on society is an enterprise well-suited for collaborations between industry and academia," Ribeiro told us. ®

More about


Send us news

Other stories you might like