A Nature study found TikTok’s algorithm systematically amplified pro‑Republican content in New York, Texas and Georgia during the 2024 U.S. presidential campaign. Researchers used 323 simulated accounts over 27 weeks and analyzed more than 280,000 recommended videos. Pro‑Republican‑trained bots saw ~11.5% more aligned content, while pro‑Democratic bots were ~7.5% likelier to be shown pro‑Republican material. The study notes limitations and calls for greater transparency and accountability for recommendation systems.
Study Finds TikTok Algorithm Tilted Toward Pro‑Republican Content Ahead Of 2024 U.S. Election
Bots in the study that were trained on pro-Republican content viewed about 11.5% more content that agreed with their views compared to their pro-Democrat counterparts.Photograph: Brent Lewin/Bloomberg via Getty Images(Photograph: Brent Lewin/Bloomberg via Getty Images)
A study published in the journal Nature reports that TikTok’s recommendation algorithm systematically prioritized pro‑Republican political content in three U.S. states during the run‑up to the 2024 presidential election.
Researchers at New York University Abu Dhabi created 323 simulated TikTok accounts and trained them to mimic real user behavior by watching curated sets of videos aligned with either the U.S. Democratic or Republican parties. The accounts were routed through mock GPS signals and VPNs to appear in New York State, Texas and Georgia. Over 27 weeks of the 2024 campaign, the team collected and analyzed more than 280,000 videos recommended to those accounts’ For You pages using a mix of human and automated review.
Key Findings
The study found measurable and consistent imbalances in what the platform recommended. Accounts trained on pro‑Republican content encountered about 11.5% more content that aligned with their views than accounts trained on pro‑Democratic content. Conversely, accounts trained on pro‑Democratic material were roughly 7.5% more likely to see pro‑Republican videos on their For You page than the reverse.
“We found a consistent imbalance,” the authors wrote in Nature. “The algorithm wasn’t just giving people what they want; it was giving one side more of what the other side says about them.”
The types of cross‑partisan content shown also differed by account type: pro‑Democratic accounts were more often fed videos attacking their side on immigration and crime, while pro‑Republican accounts received more cross‑partisan content focused on abortion. The authors suggest this pattern could indicate amplification of content designed to target an opposing side’s perceived weaknesses.
Limitations And Responses
The researchers note important limitations: the experiment captured only the early stages of a hypothetical user’s feed, analyzed English‑language transcripts (so it may miss visual or non‑English political cues), and was restricted to three states, so findings should not be generalized nationally without caution. The study does not measure whether exposure changed political beliefs or voting behavior, nor does it establish the algorithmic mechanism behind the imbalance.
TikTok responded that the experiment used fake accounts and therefore does not reflect real user behavior, adding that people in practice watch a wide variety of content and can control what they see through platform tools.
Why It Matters
The authors argue the results are relevant for discussions about platform transparency and algorithmic accountability. The Nature paper highlights regulatory contrasts: the EU’s Digital Services Act requires large platforms to assess and mitigate systemic risks to elections, while U.S. legal protections give platforms broader editorial discretion. Given that TikTok is a major source of political information for young voters, the researchers say even modest systematic differences in recommendations deserve attention.
“In an environment where margins are thin, systematic differences in the kind of political information recommended to tens of millions of young voters are worth taking seriously,” said co‑author Yasir Zaki.
Help us improve.
