TikTok disproportionately served anti-Democratic videos during the 2024 election, study finds

A recent study provides evidence that TikTok’s recommendation system tends to expose users to more conservative and anti-Democrat political content than liberal material. This ideological imbalance occurs regardless of a user’s initial political interests, suggesting that automation plays a significant role in modern information access. The research was published in the scientific journal Nature.

Researchers at New York University Abu Dhabi’s AI and Society Lab conducted the new research to explore how automated internet systems shape what political news the public sees. Experts debate whether internet polarization is driven by people seeking out their own preferred viewpoints or by computer algorithms pushing extreme content to keep users engaged.

Corresponding author Talal Rahwan noted the difficulty of answering this question on older platforms. “Our previous work studied recommendation algorithms on YouTube, and a persistent challenge was disentangling algorithmic influence from user self-selection,” he told PsyPost.

TikTok offers a helpful environment for testing algorithmic influence because its primary interface relies heavily on automation. “TikTok’s For You page, where the algorithm drives nearly everything users see, offered a uniquely clean setting to study that question, especially with the 2024 election making political content on the platform increasingly consequential,” Rahwan added.

To gather data, the researchers created 323 automated bot accounts to act as artificial users, a method sometimes called a sock puppet audit. Between April 30 and November 11, 2024, the scientists launched 21 new accounts each week. Each bot was assigned an age between 22 and 24 to mimic the habits of young adult voters.

The research team used location-masking software to virtually place these phones in specific parts of the country. They assigned the bots to New York to represent a reliably Democratic area, Texas to represent a reliably Republican area, and Georgia to act as a competitive swing state. The scientists used physical smartphones running Android software and reset the phones to factory settings after each weekly test. This process ensured the app could not track the devices across different weeks.

The accounts first went through a training phase to teach the app their supposed political tastes. Some bots watched up to 400 videos from known Republican creators, and others watched up to 400 videos from Democratic creators. The bots viewed each of these training videos for exactly one minute. In Georgia, the authors also included a set of neutral bots that skipped this political training phase entirely, acting as a control group of users with no political interest.

After the training phase, the accounts switched to the main feed to see what the algorithm would recommend. The bots watched the first ten seconds of each recommended video before moving to the next one. Across the entire 27 weeks, the researchers collected more than 280,000 recommended videos. They successfully downloaded the text transcripts for 40,264 of those videos, giving them a massive dataset of spoken content to analyze.

The authors used an ensemble of artificial intelligence language models to classify the political nature of the transcripts. They combined the outputs of three different models to prevent any single program from skewing the results. Human political science students validated the accuracy of these automated programs. The tools determined if a video was political, if it related to the upcoming election, and if it supported or opposed a specific party.

The findings suggest a distinct asymmetry in how political content is delivered on the platform. Accounts trained on Republican videos received about 11.5 percent more content aligned with their own party compared to the Democratic accounts. At the same time, the Democratic accounts were exposed to roughly 7.5 percent more cross-party content.

Co-author Yasir Zaki explained the broader implications of these numbers. “TikTok’s feed isn’t a neutral window into politics,” he said. “The platform’s recommendations treat Democrats and Republicans differently, consistently, across states, and in ways that can’t be explained by differences in how people engage with the content.”

Hazem Ibrahim, also a researcher at New York University Abu Dhabi, noted the specific nature of these results. “The skew was specifically concentrated in anti-Democratic content being pushed to Democratic-leaning accounts, rather than a generic ‘more Republican content’ effect,” Ibrahim said.

The topics of these recommended videos were also highly separated. “We were also struck by how the asymmetries clustered in particular policy domains, specifically, immigration, crime, and foreign policy for Democrats and abortion for Republicans, rather than being spread evenly across political content,” Ibrahim added.

To see if real people noticed these trends, the researchers surveyed 1,008 active United States TikTok users online. The survey asked participants if they perceived any changes in the political tone of their feeds over the past year. The online sample skewed slightly Republican and consisted mostly of adults between the ages of 25 and 34.

The survey responses provided evidence that lived experiences matched the automated bot experiment. Republican respondents were significantly more likely than Democratic respondents to report seeing positive political content that agreed with their views. When asked open-ended questions, conservative users frequently mentioned an increase in optimistic, pro-Trump posts on their daily feeds.

Rahwan highlighted the rigorous statistical testing behind the findings. “The gaps are averages across hundreds of experiments over six months that held up across three states and survived 48 robustness checks, on a platform serving political content to tens of millions of young voters daily, that consistency matters,” Rahwan said.

Robustness checks are mathematical tests used to ensure findings hold true under various conditions. Still, Rahwan noted that exposure does not equal influence. “That said, we measure exposure, not persuasion, so we can’t say this changed anyone’s vote,” he added.

Zaki also emphasized that the research does not prove the company intentionally programmed its software to favor conservatives. “We are not saying TikTok deliberately chose to favor Republicans, our study documents a pattern in outcomes, not intent,” Zaki said.

The automated accounts were specifically designed to represent new users interacting with the platform for the first time. “We also study content exposure, not attitude change, and our bots simulate new users with short engagement histories, so long-tenured users may have different experiences,” Zaki said.

The gap could also result from basic supply differences in the types of videos that political creators chose to upload during the election cycle. In addition, the study only looked at English language videos. This leaves out the experiences of Spanish speakers and other minority language communities in the United States.

The researchers plan to expand this work through their research group, which can be found online at ai-and-society.com. “We’d like to combine bot audits with real user data, develop methods to capture visual and audio political messaging beyond transcripts, and run cross-platform comparisons,” Ibrahim said. “The big open question is connecting these exposure patterns to downstream effects on attitudes and behavior.”

The study, “Systematic partisan content skews in TikTok during the 2024 US elections,” was authored by Hazem Ibrahim, HyunSeok Daniel Jang, Nouar Aldahoul, Aaron R. Kaufman, Talal Rahwan, and Yasir Zaki.