It started in the darkest corners of the internet. Through late August and early September 2024, neo-Nazis on Gab were seeding a fabricated story: Haitian immigrants in Springfield, Ohio, were eating the local pets. The claim went mainstream when a resident posted a secondhand version to a private Facebook group — something she’d heard from her neighbor’s daughter’s friend.

From there, the lie jumped to X and Truth Social. Soon, it even made its way to cable news. By the time then-presidential candidate Donald Trump repeated it at the presidential debate on Sept. 10 — before an audience of 67 million — the claim had already spawned AI-generated memes and campaign billboards. Thirty-three bomb threats followed; schools, hospitals and government buildings were evacuated.

Imagine if a fact-checker had known it was coming. Not after the damage was done, but days before it crossed from the fringes to the mainstream.

A team of USC researchers has built a system that can predict when a false rumor will jump from one social media platform to another, days before it goes mainstream. 

Their paper, “Cross-Platform Narrative Prediction: Leveraging Platform-Invariant Discourse Networks,” accepted at The Web Conference 2026, offers something the field has long lacked: a way to see across platform boundaries before a false narrative spreads. The work was led by Patrick Gerard, a USC Viterbi Ph.D. student in computer science at the USC Information Sciences Institute, and co-authored by Luca Luceri, a research assistant professor at the USC Thomas Lord Department of Computer Science and lead scientist at ISI, both within the USC Mark and Mary Stevens School of Computing and Artificial Intelligence; Leonardo Blas, a USC Viterbi Ph.D. student advised by Emilio Ferrara; and Ferrara, a principal scientist at ISI and USC Viterbi professor of computer science.

Four Platforms, Four Separate Worlds

The way researchers and platforms currently track misinformation is like watching four different traffic cameras with no insight on the streets connecting them. Each platform is monitored in isolation: hashtags on X, forwarded messages on Telegram, video views on TikTok. When a lie crosses from one platform to another, those signals vanish.

“These false narratives and rumors, they don’t stay on one platform; they jump,” Gerard said. “Something might start on Telegram, then go to Truth Social, and by the time it hits X or TikTok, it’s already reached millions of people. But most of our tools treat each platform as its own separate world. We’re missing all the connections between them that would let us see these patterns coming.”

The existing toolkit, known as diffusion models, looks for signals like shared URLs or identical hashtags to track how content moves. The problem: very few people use those signals. Only a small fraction of posts on X contain URLs. On Telegram, hashtags barely exist. These approaches, the researchers argue, fall short.

Mapping What People Talk About

The key shift was moving from mapping who people follow to examining what they talk about. It builds on a basic sociological principle: shared interests and inputs lead to similar patterns of behavior. Gerard’s contribution is recognizing that these patterns persist across platforms, rather than being confined to any single one.

To build what the researchers call a “discourse network,” the team looked for users who engaged with the same kinds of narratives over time, whether or not those users ever interacted directly. A conspiracy-minded Telegram user and a mainstream X user who both consume anti-immigrant content might never follow each other. But in a discourse network, they become neighbors.

“If user A has never interacted with user B, but they still talk about the same thing all the time, they can still be structurally related,” Gerard said. “We don’t have to rely on direct connections. We just ask: are they talking about the same stuff?”

In practice, the pipeline works in four steps. The system collects posts from multiple platforms. An AI language model strips out platform-specific slang and style, reducing each post to its core claim. Similar claims are clustered into coherent narratives. Finally, users are mapped to each other based on which narratives they engage with, creating a cross-platform web. When a narrative lights up among users on one platform, the model predicts where it will spread next.

ISI researcher Luceri called the work “very important.” He added: “In a way, we’re building a graph that puts together users from different social networks with the idea of discovering the main narratives, how they diffuse from one platform to another, and users’ participation within these narratives.”

94% Accuracy; 3% of Users.

To test the approach, the team analyzed 5.7 million posts from X, TikTok, Truth Social and Telegram, collected between April and November 2024, covering the arc of the U.S. presidential election. Nearly one million distinct users were represented across text posts, video transcripts and forwarded messages.

The results were striking. The discourse network model achieved over 94% accuracy, outperforming diffusion models by 55% and besting the next-best competing approach by 27%. Prediction windows of three, seven and 14 days all held up. Even more remarkable: the system needed data from only 2.9% of active users to maintain 97% of its peak performance.

Applied retrospectively to Springfield, the system flagged the narrative on Telegram on Sept. 2, 2024, and correctly forecast it would emerge on X within three days. It did. The claim appeared on X on Sept. 5, reached 1,100 posts by Sept. 6 and surged past 9,000 by Sept. 7. A similar analysis of the Hurricane Helene FEMA conspiracy, a false claim that federal workers were blocking private relief efforts, correctly predicted the jump to X within seven days of detecting it on Telegram.

Lead Time Changes Everything

The researchers are careful not to oversell what this system can do on its own. Predicting a false narrative is about to go mainstream does not lead platforms to remove it. But it gives journalists, fact-checkers and civil society organizations something they rarely have: time.

“A three-day window or a seven-day window, however much lead time we can give fact-checking organizations and journalists to debunk things that are about to spread to a more mainstream platform, that’s going to help us tremendously,” Gerard said. “We can’t stop it entirely, but we can at least be in the room when it arrives.”

Organizations such as PolitiFact, Snopes, AP Fact Check and the BBC [Verify] could use early warnings to prepare rebuttals before a story explodes, rather than scrambling after millions have already seen it, Gerard said.

Added paper co-author Ferrara: “Policymakers and crisis responders get lead time during fast-moving events such as crises, and society benefits when interventions land before narratives reach mass amplification.”

Gerard is clear-eyed about the limits. Social media platforms have little economic incentive to act on this kind of intelligence. Engagement drives revenue, and harmful narratives are often highly engaging. “The platform incentives are fundamentally misaligned with information quality,” he said. “We can build these tools, we can lead the horse to water, but if the horse’s business model depends on not drinking, this is not going to be enough.”

Gerard is now working with co-authors on a study of how podcast ecosystems correlated with social media behavior during the 2024 election. He also is pursuing research to extend the discourse network approach to as many as 15 or 20 platforms. 

The stakes, Gerard believes, are huge. False narratives thrive when they move faster than the truth. A system that narrows that gap, even modestly, serves something larger than any single debunking.

“Democratic society relies on the coexistence, in the same world, of facts,” Gerard said. “We have to lay that groundwork to stay within the same reality.”

Published on June 24th, 2026

Last updated on June 24th, 2026