Can AI talk us out of conspiracy theory rabbit holes?

Dana McKay and Johanne Trippas for The Conversation

Melbourne: New research published in Science shows that for some people who believe in conspiracy theories, a fact-based conversation with an artificial intelligence (AI) chatbot can “pull them out of the rabbit hole”.

Better yet, it seems to keep them out for at least two months.

This research, carried out by Thomas Costello at the Massachusetts Institute of Technology and colleagues, shows promise for a challenging social problem: belief in conspiracy theories.

Some conspiracy theories are relatively harmless, such as believing Finland doesn’t exist (which is fine, until you meet a Finn). Other theories, though, reduce trust in public institutions and science.

This becomes a problem when conspiracy theories persuade people not to get vaccinated or not to take action against climate change. At its most extreme, belief in conspiracy theories has been associated with people dying.

Conspiracy theories are ‘sticky’ Despite the negative impacts of conspiracy theories, they have proven very “sticky”. Once people believe in a conspiracy theory, changing their mind is hard.

The reasons for this are complex. Conspiracy theorist beliefs are associated with communities, and conspiracy theorists have often done extensive research to reach their position.

When a person no longer trusts science or anyone outside their community, it’s hard to change their beliefs.

Enter AI The explosion of generative AI into the public sphere has increased concerns about people believing in things that aren’t true. AI makes it very easy to create believable fake content.

Even if used in good faith, AI systems can get facts wrong. (ChatGPT and other chatbots even warn users that they might be wrong about some topics.) AI systems also contain widespread biases, meaning they can promote negative beliefs about some groups of people.

Given all this, it’s quite surprising that a chat with a system known to produce fake news can convince some people to abandon conspiracy theories, and that the change seems to be long lasting.

However, this new research leaves us with a good-news/bad-news problem.

It’s great we’ve identified something that has some effect on conspiracy theorist beliefs! But if AI chatbots are good at talking people out of sticky, anti-scientific beliefs, what does that mean for true beliefs? What can the chatbots do? Let’s dig into the new research in more detail. The researchers were interested to know whether factual arguments could be used to persuade people against conspiracy theorist beliefs.

This research used over 2,000 participants across two studies, all chatting with an AI chatbot after describing a conspiracy theory they believed. All participants were told they were talking to an AI chatbot.

The people in the “treatment” group (60 per cent of all participants) conversed with a chatbot that was personalised to their particular conspiracy theory, and the reasons why they believed in it.

This chatbot tried to convince these participants that their beliefs were wrong using factual arguments over three rounds of conversation (the participant and the chatbot each taking a turn to talk is a round). The other half of participants had a general discussion with a chatbot.

The researchers found that about 20 per cent of participants in the treatment group showed a reduced belief in conspiracy theories after their discussion. When the researchers checked in with participants two months later, most of these people still showed reduced belief in conspiracy theories. The scientists even checked whether the AI chatbots were accurate, and they (mostly) were.

We can see that for some people at least, a three-round conversation with a chatbot can persuade them against a conspiracy theory.

So we can fix things with chatbots? Chatbots do offer some promise with two of the challenges in addressing false beliefs.

Because they are computers, they are not perceived as having an “agenda”, making what they say more trustworthy (especially to someone who has lost faith in public institutions).

Chatbots can also put together an argument, which is better than facts alone. A simple recitation of facts is only minimally effective against fake beliefs.

Chatbots aren’t a cure-all though. This study showed they were more effective for people who didn’t have strong personal reasons for believing in a conspiracy theory, meaning they probably won’t help people for whom conspiracy is community.

So should I use ChatGPT to check my facts? This study demonstrates how persuasive chatbots can be. This is great when they are primed to convince people of facts, but what if they aren’t? One major way chatbots can promote misinformation or conspiracies is when their underlying data is wrong or biased: the chatbot will reflect this.

Some chatbots are designed to deliberately reflect biases or increase or limit transparency. You can even chat to versions of ChatGPT customised to argue that Earth is flat.

A second, more worrying probability, is that as chatbots respond to biased prompts (that searchers may not realise are biased), they may perpetuate misinformation (including conspiracy beliefs).

We already know that people are bad at fact checking and when they use search engines to do so, those search engines respond to their (unwittingly biased) search terms, reinforcing beliefs in misinformation. Chatbots are likely to be the same.

Ultimately, chatbots are a tool. They may be helpful in debunking conspiracy theories – but like any tool, the skill and intention of the toolmaker and user matter. Conspiracy theories start with people, and it will be people that end them.

(Dana McKay is Associate Dean, Interaction, Technology and Information at RMIT University and Johanne Trippas is a Research fellow at RMIT University.)