Source-linked AI summary
Dynamics of Algorithmic Content Amplification on TikTok
Fabian Baumann, Nipun Arora, Iyad Rahwan, Agnieszka Czaplicka
TL;DR
The paper investigates how quickly and extensively TikTok’s For You algorithm amplifies content aligned with users’ interests, an area that remains largely unquantified. Using a sock-puppet audit with interest-specific bots and longitudinal recommendation analysis, it finds rapid, sustained amplification and examines its relationship with content diversity. The study also acknowledges limitations of using controlled bots to represent user behavior.
Problem
The paper addresses the largely unclear dynamics and extent of TikTok’s amplification of content aligned with users’ interests.
Method
The study deploys automated bots with distinct content preferences to interact with TikTok’s For You feed, combining a sock-puppet audit with time series and Markov-model analyses.
Results
TikTok’s algorithm strongly amplifies interest-aligned content, with rapid reinforcement typically occurring within approximately 200 videos.
Takeaways & Limitations
The findings highlight a trade-off between personalized reinforcement and exposure to new topics and perspectives in TikTok’s recommendation feed.
Takeaways & Limitations
The study acknowledges limitations associated with using controlled bots to investigate recommendation behavior.
Abstract
from arXiv · showhide
Intelligent algorithms increasingly shape the content we encounter and engage with online. TikTok's For You feed exemplifies extreme algorithm-driven curation, tailoring the stream of video content almost exclusively based on users' explicit and implicit interactions with the platform. Despite growing attention, the dynamics of content amplification on TikTok remain largely unquantified. How quickly, and to what extent, does TikTok's algorithm amplify content aligned with users' interests? To address these questions, we conduct a sock-puppet audit, deploying bots with different interests to engage with TikTok's "For You" feed. Our findings reveal that content aligned with the bots' interests undergoes strong amplification, with rapid reinforcement typically occurring within the first 200 videos watched. While amplification is consistently observed across all interests, its intensity varies by interest, indicating the emergence of topic-specific biases. Time series analyses and Markov models uncover distinct phases of recommendation dynamics, including persistent content reinforcement and a gradual decline in content diversity over time. Although TikTok's algorithm preserves some content diversity, we find a strong negative correlation between amplification and exploration: as the amplification of interest-aligned content increases, engagement with unseen hashtags declines. These findings contribute to discussions on socio-algorithmic feedback loops in the digital age and the trade-offs between personalization and content diversity.
1 Introduction
TikTok’s algorithm-driven For You feed rapidly adapts content exposure to users’ explicit and implicit interactions, but the extent of interest-aligned amplification remains unclear. The paper addresses this gap with a sock-puppet audit that measures how recommendation dynamics vary across interests and over time.
- 1 Introduction: TikTok’s For You feed dictates content exposure primarily through users’ explicit and implicit interactions with previously encountered videos.These interactions include rewatching, liking, and following creators, forming a direct feedback loop through which the algorithm adapts to preferences.
- 1 Introduction: The study asks how quickly and to what extent TikTok amplifies content aligned with users’ interests.Prior work examined several aspects of TikTok’s feed, but the distributional effects of interest-aligned amplification remained largely unclear.
- 1 Introduction: A sock-puppet audit deploys automated bots with distinct content preferences to interact in real time with TikTok’s For You feed.The approach extends previous sock-puppet methodologies by using GPT-3.5 Turbo to assess video relevance from descriptions and hashtags, then rewatching, liking, and following aligned content.
- 1 Introduction: The experiments compare bots interested in GAMING, FOOD, or both GAMING and FOOD, using prevalent topic categories to enable interest signaling within a reasonable timeframe.Bots classify encountered videos as GAMING, FOOD, or both and interact when content matches their assigned interests.
- 1 Introduction: The analysis examines detection speed, changing amplification across interests, relationships with content characteristics, and whether reinforcement narrows exposure to new topics and perspectives.Time series analysis and (hidden) Markov models are used to identify recommendation patterns, while amplification is related to popularity, duration, and content diversity.
2 Results
TikTok rapidly amplifies interest-aligned recommendations across single- and dual-interest bots, with amplification strength and dynamics varying by topic. The process includes persistent reinforcement, topic-dependent transitions, hidden-state complexity, and reduced exploration as alignment increases.
- Content Amplification: 67.4%, 52.3%, and 67.2% of videos were interest-aligned for GAMING, FOOD, and GAMING+FOOD bots, respectively, exceeding both baseline expectations across conditions.Amplification was strongest for GAMING, lower for FOOD, and more moderate for dual-interest bots.
- Content Amplification: Most amplification onset points occurred within the first 200 videos, averaging 65.7 videos for GAMING, 140 for FOOD, and 93.2 for GAMING+FOOD bots.The onset marks a sharp initial increase in the rate of interest-aligned content.
- Markov Model: Markov dynamics showed stronger persistence for GAMING than FOOD, with GAMING stationary interest-aligned content at 67.9% versus FOOD at 53.4%.GAMING content followed non-interest content 67% of the time and interest-aligned content 68.4% of the time; FOOD aligned content followed non-interest content 49.7% of the time and aligned content 56% of the time.
- Hidden Markov Model: The Hidden Markov Model inferred the greatest dynamical complexity for FOOD, which had more hidden states and substantially more state transitions than GAMING or GAMING+FOOD.GAMING examples were best described by one hidden state, whereas FOOD and GAMING+FOOD examples were best described by two.
- Content Exploration: Exploration and amplification were strongly negatively correlated, strongest for GAMING (r = −0.92) and weakest for GAMING+FOOD (r = −0.71).FOOD bots explored more unique hashtags than GAMING bots, while dual-interest bots generally fell between the single-interest conditions.
3 Discussion
TikTok’s For You feed rapidly and persistently amplifies interest-aligned content, with amplification varying across topics and recommendation dynamics unfolding in distinct phases. This personalization is accompanied by reduced exploration, although the feed preserves some content diversity.
- 3 Discussion: Interest-aligned content shows elevated transition probabilities across conditions, while Hidden Markov Models identify interest-specific differences in hidden states and transitions.These analyses formalize distinct recommendation dynamics after the initial amplification phase.
- 3 Discussion: Within approximately 200 videos, TikTok rapidly amplifies interest-aligned content, with GAMING bots showing the fastest onset and FOOD bots the slowest.The paper equates this period to roughly 1.5 hours of browsing.
- 3 Discussion: Amplification varies by topic: GAMING content undergoes greater amplification than FOOD content, indicating topic-specific biases in TikTok’s feed algorithm.Interest-aligned content is also less popular by like counts and longer in duration than non-interest content, suggesting a trade-off between personalization and mainstream appeal.
- 3 Discussion: Even in the strongest amplification cases, roughly one-third of content remains unaligned with users’ interests, indicating that TikTok preserves exploration rather than complete convergence.Similar stationary probabilities for single- and dual-interest bots indicate that fewer interests do not necessarily produce more aggressive amplification.
- 3 Discussion: More interest-aligned content is strongly associated with less exploration, measured by fewer unique hashtags encountered and lower individual-level content diversity.The findings describe a potential downside of algorithmic amplification: reinforcement can restrict exposure to new content.
- 3 Discussion: The sock-puppet audit provides controlled insights but is limited by non-human engagement patterns, metadata-based topic inference, the absence of a neutral baseline, and focus on GAMING and FOOD bots in the US.Longer studies and broader topics may be needed to assess whether amplification plateaus, escalates, or oscillates and whether the findings generalize globally.
4 Methods
The study uses a sock-puppet audit with interest-specific TikTok bots that mimic user interactions under controlled browsing conditions. Across seven test cycles, 42 bots processed 128,148 videos, producing a cleaned dataset of 112,610 videos for analysis.
- Automation and controls: The audit automated TikTok browsing with Selenium and Selenium Stealth, American-server proxies, saved cookies, and randomized inactivity periods to approximate human-like activity.Distinct Chrome profiles enabled automatic re-authentication and helped circumvent CAPTCHA verification.
- Data collection: 42 bots completed seven test cycles, processing 128,148 videos and yielding 112,610 videos after excluding entries without descriptions or hashtags.The final dataset included 58,213 unique entries, 817,200 hashtags, and 27,507 distinct creators.
- Data processing: The study collected video metadata including duration, descriptions, hashtags, creator names, and likes in MongoDB Atlas for subsequent analysis.The average video duration was 79 seconds, and bots watched approximately 64% of each video's length on average.
- Scope and reproducibility: TikTok updates and unmaintained code may require modifications for future replication of the auditing framework.The authors caution that platform changes can affect framework functionality and compatibility.
Supplemental Material of Dynamics of Algorithmic Content Amplification on TikTok
Supplemental analyses use Markov and Hidden Markov models, change-point detection, and curve fitting to characterize TikTok recommendation dynamics. These methods quantify amplification onset and trends while examining transitions, hidden states, and variation across experimental conditions.
- Markov models: Markov transition matrices are computed from observed state transitions and normalized row-wise to represent probabilities, with stationary distributions derived from eigenvectors for eigenvalue 1.The models support analysis of long-run recommendation-state distributions.
- Hidden Markov Models: Hidden Markov Models analyze recommendation dynamics beyond observable state transitions using binary signals derived from bot interactions with TikTok content.Model selection compares different numbers of hidden states and random initializations using Bayesian Information Criterion scores.
- Experimental comparisons: The supplemental material summarizes bot-level statistics, including videos watched and the proportion of interest-aligned content, to compare amplification patterns across experiments.Tables 1–3 report statistics for the GAMING, FOOD, and GAMING+FOOD conditions.
- Amplification dynamics: Change-point detection identifies when TikTok's feed begins significantly reinforcing interest-aligned content, while curve fitting estimates its rate of increase over time.These analyses provide more precise quantification of amplification dynamics across experimental conditions.
- Experimental conditions: The supplemental figures present GAMING, FOOD, and GAMING+FOOD experimental conditions for examining recommendation dynamics.Figures 11–13 correspond respectively to the three conditions.