Researchers Use AI to Separate Conspiracy From Fact

Researchers at UCLA have created a model to distinguish conspiracy theories from actual conspiracies which will possibly allow robots to differentiate between things like Pizzagate and Bridgegate in the future.

The project, created by Timothy R. Tangherlini, Shadi Shahsavari, Behnam Shahbazi, Ehsan Ebrahimzadeh, and Vwani Roychowdhury, analyzes multiple narratives including “user-generated” narratives like Pizzagate, Qanon, the Illuminati, etc., and compares them with real narratives that seemed like conspiracies.

The process breaks stories up into separate chunks and assesses the actants—the humans, objects, or places involved in a story. The narrative can be a series of posts in a forum or blog or a bunch of Twitter data.

The researchers found that conspiracy theory narratives are often choppy and easily picked apart because the number of actants is smaller and each one is more important. Further, conspiracy theories often lose coherence as information is added while actual conspiracies become stronger.

“Reporting on actual conspiracies introduces new actants and relationships as part of the process of validating what has actually happened,” wrote the researchers. “This reporting feeds the core giant network with more evidence, resulting in a denser network over time. Conspiracy theories, by way of contrast, may form rapidly. Since the only evidence to support any of the actants and relationships comes from the storytellers themselves, we suggest that the network structure of a conspiracy theory stabilizes quickly. This stabilization is supported by studies in folklore, which reveal that an essentially constant and relatively small set of actants and relationships determines the boundaries of admissible stories (or story fragments) after the initial narrative burst finishes.”

In other words, once a conspiracy theory forms, it rapidly focuses on a few major figures—Trump, JFK Jr., a lizard monster—but never becomes more complex. Actual conspiracies grow and become more informationally dense over time. Imagine, for example, Cinderella vs. the story of Helen Keller (who, incidentally, is being attacked by TikTokkers as being a fraud.) While Cinderella has few confirmations and a limited set of actants, the Helen Keller story involves many people over a long timeline and as information is added, be it documents, images, or videos, the story becomes clearer.

Reality, it seems, is denser than fiction.

“So, a real narrative when subjected to this process, tends to create a network that is hard to break up into smaller connected parts—with Bridgegate or with Enron, everyone is part of some other preexisting community (and they are all from that community), so deleting people and their relationships does not fracture or split the narrative graph,” said Tangherlini. “For something like a conspiracy theory—a fictional account—it is very easy to fracture the graph since one or two actants and/or relationships provide the glue holding otherwise unconnected parts of the framework together. So, for instance, unless you as a storyteller have access to secret or esoteric knowledge allowing you to find “hidden” (tenuous) connections, there is no immediate reason to connect 5G to Bill Gates to Covid-19 to Global warming. These connections are “discovered” by the conspiracy theorists, and become the weak links in their narrative frameworks.”

Conspiracy theories can get denser over time, but not in the same way.

“Unfortunately, it is very easy to create such links (just as it is very easy to show them to be spurious),” he said. “Once there are enough people activating those links in their storytelling, they gain weight in the narrative framework.”

The researchers believe that forums and social media companies can use this to warn users about potential fake news.

“The pipeline has a great degree of AI baked-in since there are quite a number of classification tasks that have to be accomplished to jointly estimate all the parameters of the system. Right now, the pipeline is largely automated,” said Tangherlini. “There are a couple of choke points: data acquisition and cleaning are challenging for a number of reasons; we don’t have the computing resources to do the real-time analysis so we do human-interactive time; there is some human input that occurs to choose best labels for nodes/edges.”

The team is working on a way to automate the process of truth-finding, a dream that many of us share as we peruse the internet these days.