Identifying Political Manipulation Behind TikTok's Algorithm

How to cite

Chen, Kuan-Ming; Chu, C. Y. Cyrus; Lin, Ming-Jen; Chen, Ming-Chi; Lai, Chien-Yu; Chen, Yu-Chang; Sheen, Greg Chih-Hsin; Wang, Chan-Hsi. 2026. "Identifying Political Manipulation Behind TikTok's Algorithm." AEA RCT Registry. AEARCTR-0018374. https://doi.org/10.1257/rct.18374-1.0

Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.

Kuan-Ming Chen* , C. Y. Cyrus Chu , Ming-Jen Lin , Ming-Chi Chen , Chien-Yu Lai , Yu-Chang Chen , Greg Chih-Hsin Sheen , Chan-Hsi Wang

* Primary investigator

Registry ID
AEARCTR-0018374
DOI
https://doi.org/10.1257/rct.18374-1.0
First registered
Last updated
Status
In development
Trial dates
2026-05-01 – 2026-12-31
Countries
Private
Registry record
http://www.socialscienceregistry.org/trials/18374

Abstract

This study investigates whether TikTok's recommendation algorithm systematically suppresses content that is politically sensitive to the Chinese Communist Party (CCP). In the current trial, we use 200 mobile phones, each configured with a unique TikTok account, to upload 13 distinct text-based video stimuli—each uploaded 100 times across devices—for a total of 1,300 uploads. The stimuli are paired treatment/control texts (in Chinese) spanning five thematic domains: (1) "Wuhan Pneumonia" (武漢肺炎) vs. "COVID-19," (2) "Falun Gong" (法輪功) vs. "Tai Chi" (太極拳), (3) "Support Taiwan Independence" (支持台灣獨立) vs. "Taiwan Independence" (台灣獨立) vs. "Oppose Taiwan Independence" (反對台灣獨立), (4) "Support Palestine Independence" (支持巴勒斯坦獨立) vs. "Palestine Independence" (巴勒斯坦獨立) vs. "Oppose Palestine Independence" (反對巴勒斯坦獨立), and (5) "East Turkistan" vs. "West Turkistan" vs. "Xinjiang." Sensitive text is embedded in TikTok posts using a graphic overlay technique so that only TikTok's algorithm—not human viewers—can identify the text content. The primary outcome is the play count of each video measured at multiple intervals after upload (up to 12 hours).

Experimental design

The whole experimental design is described in the hidden part.

Design details

Randomization method
Randomization is done by a computer.
Randomization unit
Each of the 13 video scripts is uploaded approximately 100 times, yielding 1,300 total uploads. 200 mobile phones are used; each phone uploads one video per half-day session.
Clusters
N/A (not clustered)
Observations
1,300 (13 video scripts × 100 uploads each)
Arms
Each of the 13 video scripts is uploaded approximately 100 times, yielding 1,300 total uploads. For a detailed breakdown of the unit, see the experimental design.
Primary outcomes
Play count (views) at 12 hours post-upload: The number of views each uploaded video receives within 12 hours of posting, collected via automated web scraping.
Secondary outcomes
Play count at intermediate intervals: Views measured at 0.1, 0.5, 1, 3, and 6 hours after upload. Like count at 12 hours post-upload: The number of likes each video receives within 12 hours. Comment count at 12 hours post-upload: The number of comments each video receives within 12 hours.

Cited by

No citation data available for this registration yet.

Keywords