For example, 9/11 deniers often point to the claim that jet fuel doesn’t burn hot enough to melt steel as evidence that airplanes were not responsible for bringing down the Twin Towers—but the chatbot responds by pointing out that although this is true, the American Institute of Steel Construction says jet fuel does burn hot enough to reduce the strength of steel by over 50%, which is more than enough to cause such towers to collapse.
Although we have greater access to factual information than ever before, it is extremely difficult to search that vast corpus of knowledge efficiently. Finding the truth that way requires knowing what to google—or who to listen to—and being sufficiently motivated to seek out conflicting information. There are large time and skill barriers to conducting such a search every time we hear a new claim, and so it’s easy to take conspiratorial content you stumble upon at face value. And most would-be debunkers at the Thanksgiving table make elementary mistakes that AI avoids: Do you know the melting point and tensile strength of steel offhand? And when your relative calls you an idiot while trying to correct you, are you able to maintain your composure?
With enough effort, humans would almost certainly be able to research and deliver facts like the AI in our experiments. And in a follow-up experiment, we found that the AI debunking was just as effective if we told participants they were talking to an expert rather than an AI. So it’s not that the debunking effect is AI-specific. Generally speaking, facts and evidence delivered by humans would also work. But it would require a lot of time and concentration for a human to come up with those facts. Generative AI can do the cognitive labor of fact-checking and rebutting conspiracy claims much more efficiently.
In another large follow-up experiment, we found that what drove the debunking effect was specifically the facts and evidence the model provided: Factors like letting people know the chatbot was going to try to talk them out of their beliefs didn’t reduce its efficacy, whereas telling the model to try to persuade its chat partner without using facts and evidence totally eliminated the effect.