AutoLike: Auditing Social Media Recommendations through User Interactions

Athina Markopoulou; Hieu Le; Salma Elmalaki; Zubair Shafiq

REVIEW 4 major objections 5 minor 1 cited by

AutoLike: Auditing Social Media Recommendations through User Interactions

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read AutoLike, a reinforcement-learning agent that interacts with a recommendation system through ordinary user actions, can drive TikTok to serve several times more of a chosen topic and sentiment—including sad mental-health content—starting…

desk verdict A sensible RL framing for auditing TikTok, but the headline 2–3x 'Sad Mental Health' effect is measured with the same classifier that decides the likes, and the deployed experiments never actually run the RL policy. read the letter →

arxiv 2502.08933 v1 pith:5CQULPAV submitted 2025-02-13 cs.LG

Hieu Le , Salma Elmalaki , Zubair Shafiq , Athina Markopoulou This is my paper · ORCID

classification cs.LG

keywords algorithmicauditingrecommendationsystemsTikTokreinforcementlearningzero-shotclassificationmentalhealthonlineharmssocialmediapersonalization

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper is trying to establish that a recommendation algorithm can be audited from the outside by an automated agent that only uses the same interactions a normal user has—liking, skipping, watching. The agent treats the 'For You' page as a black-box environment and learns which actions push it toward a chosen target, such as negative mental-health content. The authors demonstrate the idea on TikTok with a streamlined version: liking videos that a zero-shot classifier judges to be on-topic and sad, skipping everything else, then counting what the feed serves. The headline result is that this simple interaction pattern drove TikTok to serve roughly twice as much sad mental-health content as a skip-only control feed, and up to three times as much by the end of the run. If the method works, it gives regulators and researchers a way to test what platforms deliver without relying on platform cooperation or internal data.

What carries the argument

The load-bearing object is the reinforcement-learning loop that turns a black-box recommendation feed into a measurable state. A state is a pair $\langle\text{topic}, \text{sentiment}\rangle$, each a confidence score in [0,1] binned to tenths; the reward is $1 - \frac{\text{distance}(s_{t+1}, g)}{\sqrt{2}}$ for a target $g$. The agent updates state-action values with Q-learning and chooses actions (like or skip) according to a policy—in the streamlined deployment, it likes when the classifier's confidence exceeds a threshold (0.5 for single-dimension runs, 0.6 for the topic-plus-sentiment runs) and skips otherwise. The state itself is computed by automatically extracting each TikTok's description and transcribing its audio, concatenating them, and passing the text through a zero-shot natural-language-inference classifier that returns on-topic and negative-sentiment confidences. This machinery matters because it lets the agent learn the effect of its own actions on a system whose algorithm is completely unknown, and it converts 'which content appears' into a numerical reward the agent can optimize.

What would settle it

Manually label a sample of the videos from the live 'Sad Mental Health' experiment (AUTOLIKE-D2) and recompute the on-topic and sad counts using human judgments instead of classifier confidence; if the 2×–3× gap over the skip-only control shrinks to near zero under human labels, the effect is an artifact of the classifier rather than a change in TikTok's behavior.

Watch

Extended reading notes

Core claim

The paper claims that a recommendation system can be steered, not just observed, by a learning agent acting through the platform's own user interface. The agent's state is the current video represented as a pair of confidence scores in [0,1]—how on-topic it is and how negative its sentiment is—and the reward is the normalized proximity of that state to a target chosen by the examiner (e.g., topic = mental health, sentiment = sad). Using Q-learning over the action space of liking versus skipping, the streamlined agent learns from experience which actions move the feed toward the target. Across eight live experiments on fresh TikTok accounts, the topic-and-sentiment runs served 1.5–2× as much matching content as the skip-only control, with the 'Sad Mental Health' run reaching 2× at step 125 and 3× by the end of the ~150-step pathway. The paper also shows that driving sentiment alone, without a topic, produced no discernible shift, which it interprets as TikTok's recommender prioritizing topic personalization over sentiment.

Load-bearing premise

The live experiments measure success with the same zero-shot classifier that decides which videos to like, so if that classifier over-labels certain videos as sad mental-health content, the reported 2×–3× effect could be a classification artifact rather than a real change in TikTok's recommendations.

Editorial extensions

If this is right

Regulators and platform developers could deploy AutoLike to measure whether a platform serves harmful content, and how few interactions are needed to reach it.
The framework works from a fresh account with no explicit interests, meaning purely interaction-based signals are sufficient to personalize a feed toward a chosen topic.
The sentiment-only result suggests that platforms may weight topic identity above sentiment, so attempts to avoid negative content may need to target topic rather than mood.
AutoLike is agnostic to the content classifier, so its auditing capability improves automatically as classification models improve.
The same setup can be adapted to other 'For You'-style platforms (e.g., Instagram Reels, YouTube Shorts) and to other content dimensions such as truthfulness or intent.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

An unresolved concern is that the same zero-shot classifier both selects the videos to like and counts the outcome; if that classifier over-labels certain audio or description patterns as 'sad mental health,' the measured 2×–3× could partly be a measurement artifact rather than a genuine change in TikTok's recommendations.
If the steering effect is real, it suggests that a small number of ordinary actions can outweigh a platform's safety filters, which would be a concrete exploit for users who want to avoid or seek particular content.
A natural next experiment would use the same framework but with human-labeled outcome labels on the live feeds, or with a hold-out classifier different from the one that drives the actions, to separate personalization from classifier bias.
Beyond topics and sentiment, the same RL audit could measure how easy it is to push a feed toward misinformation or political radicalization, turning AutoLike into a general-purpose measurement instrument for recommendation platforms.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

First thing to know: if you read one section, read Sec. 4.2. The paper's real claim is there, and it is both more fragile and more interesting than the abstract suggests. AutoLike is a clean RL framing for black-box auditing — define state as (topic, sentiment), reward as proximity to an end goal, and let the agent learn which likes/skips move the feed. That framing is genuinely new; prior sock-puppet audits mostly used hardcoded rules. The two-dimensional topic+sentiment goal also produces a useful null result: sentiment-only steering doesn't shift TikTok's feed, which hints the recommender weights topic over sentiment. The implementation is concrete and honestly reported, including the flaky parts.

The soft spot is the one you'd expect from the framing: the success metric is the same BART-large-MNLI confidence score that decides whether to like. In Fig. 6, the y-axis is literally 'Classification > 0.6' — the same threshold that triggers the like. So the reported 2x–3x increase in 'Sad Mental Health' could be the agent amplifying the classifier's own false positives, not a real change in TikTok's recommendations. Manual verification on D1 helps for the offline hashtag set, but the live D2 feeds were not hand-labeled. That's the load-bearing issue, and it's fixable: sample some of the liked TikToks from each condition and have a human label them, or at least use a different classifier for the metric.

Second issue: the paper calls it a reinforcement-learning framework, but the deployed 'streamlined' version replaces the RL agent with a fixed threshold rule — like if confidence > 0.6, no Q-learning, no exploration. The authors are upfront about this, but it means the RL part is untested. The demonstrated mechanism is a hand-set heuristic. That's fine as a proof-of-concept, but the title and abstract overstate it.

Third, smaller: single run per condition, no error bars, no code/data yet. For an auditing case study that's not disqualifying, but it does make the 2x–3x number exactly as strong as the single run.

Overall: for someone building an algorithmic auditor, this is a useful blueprint and the negative sentiment result is worth knowing. The circularity concern is serious enough that the headline empirical claim shouldn't be taken at face value. I'd send it to review, but with a clear request to address the metric/label issue and either run the actual RL policy or reframe the contribution. If that happens, this could be a solid paper; right now it's a promising one with a central number that needs independent confirmation.

Referee Report

4 major / 5 minor

Summary. The paper introduces AutoLike, a framework that casts recommendation-system auditing as a reinforcement-learning problem and applies it to TikTok. The agent takes user actions (e.g., like, skip) on a fresh For You page, using a zero-shot classifier (BART-large-MNLI) on concatenated video description and Whisper transcription to compute a (topic, sentiment) state and a proximity-based reward. The paper evaluates the classifier on a hashtag-collected dataset (AUTOLIKE-D1) and then runs eight live experiments (AUTOLIKE-D2) with a streamlined policy that likes a video when the classifier confidence exceeds 0.5 or 0.6. The central reported result is that the "Sad Mental Health" experiment yields approximately 2x (rising to 3x) as much sad mental-health content as a skip-only controlled experiment (Sec. 4.2, Figs. 5-6).

Significance. If the central claim holds, AutoLike would be a practical, low-cost tool for regulators and researchers to probe recommendation systems for harmful content without platform cooperation. The paper's strengths include a realistic Android deployment with genuine TikTok accounts, a clearly described pipeline, and an explicit comparison to prior sock-puppet approaches. The framing of auditing as RL is conceptually interesting and the authors are transparent about deployment limitations and ethical choices. However, the current evidence is undermined by two load-bearing issues: the outcome metric is computed with the same classifier that drives the actions, and the experiments do not actually exercise the RL component described in Sec. 2. The manual verification of the classifier is limited to the offline dataset, not the live feeds, and each experimental condition is run only once. These issues must be addressed before the claimed 2x-3x effect can be accepted as a real shift in TikTok's recommendations rather than a classifier artifact or a statistical fluke.

major comments (4)

[Sec. 4.2, Figs. 5 and 6; Sec. 3.2.2] The success metric is not independent of the action selection mechanism. The y-axis of Figs. 5 and 6 counts videos that the BART-large-MNLI classifier labels as on-topic/on-sentiment with confidence above 0.5 or 0.6, and the same classifier confidence at the same thresholds is what triggers the like action in the streamlined policy. Consequently, if the classifier systematically false-positives on some pattern present in live TikTok feeds (e.g., sad audio, certain creators, or phrasing not removed by the hashtag filter), the agent will preferentially like those false positives, TikTok will serve more such items in response, and the metric will count the same false positives as evidence of success. The manual verification in Sec. 4.1 (Table 3) is performed on the hashtag-collected dataset AUTOLIKE-D1, not on the live experimental feeds AUTOLIKE-D2, so the classifier's precision and recall on the actual test distribution are unknown. The paper should manually label a sample of the D2 feeds, report the classifier's performance on that sample, and ideally use an outcome metric that is not derived from the same model that selects actions (e.g., human annotation, a different classifier, or a content-based feature that is not used for reward computation).
[Sec. 4.2 vs. Sec. 2] The deployed "streamlined AutoLike" is not the RL agent defined in Sec. 2. The abstract and Sec. 1 describe a reinforcement-learning framework that "learns over time which actions most efficiently drive the algorithm," and Sec. 2 formalizes Q-learning with state-action values (Eq. 2). However, the experiments in Sec. 4.2 replace the policy with a fixed threshold: like if the classifier confidence exceeds 0.5 (or 0.6 in the multi-dimensional experiments). There is no Q-learning, no exploration/exploitation, no temporal-difference update, and no learning over actions. As a result, the empirical demonstration does not validate the RL formulation that is the paper's main conceptual contribution. The authors should either implement and evaluate the actual RL policy (e.g., Q-learning with the reward of Eq. 1) or reframe the paper's claims to present the threshold policy as a heuristic control rule and position RL as a proposed extension rather than an evaluated component.
[Sec. 4.2, Figs. 5 and 6] The central empirical claims rest on a single run per condition. Each experiment uses one fresh TikTok account and one trajectory, and the reported 2x/3x increases are comparisons between single treated runs and a single skip-only control. No error bars, confidence intervals, or statistical tests are provided. Given that TikTok's FYP is stochastic and the classifier has nontrivial error rates (Table 3 reports F1 scores as low as 0.49 for audio-only positive sentiment), the observed differences could arise from random variation in the initial FYP state. The paper should replicate each condition (or at least the control and the "Sad Mental Health" condition) and report a statistical comparison, such as confidence intervals on the cumulative counts or a permutation test, so that the reader can assess whether the effect is reproducible.
[Sec. 4.2, "Streamlining AutoLike"] The streamlined policy introduces several free parameters that are not systematically varied or justified: the classification confidence threshold (0.5 vs. 0.6), the stopping condition (at least 4 of the last 10 TikToks liked), the time horizon (100 training videos + 50 test videos), and the state discretization step (0.1). The switch from 0.5 to 0.6 is explained only as "to reduce noise," but no sensitivity analysis is provided. Since the paper's main evidence comes from this streamlined policy, the dependence of the 2x-3x result on these parameter choices should be examined, or at least the sensitivity should be discussed explicitly.

minor comments (5)

[Sec. 4.1, Fig. 4] The text says the classifier is evaluated across 24 hashtags, but Fig. 4 shows results for only four hashtags (#mentalhealth, #hatespeech, #ed, #foryou); the remaining hashtags are mentioned in the caption but not visualized, which makes it difficult to assess the full evaluation.
[Table 3] The manual verification dataset is small (63 mental-health and 63 other TikToks) and covers only mental health, not the other eight topics used in the live experiments; the authors should acknowledge this limitation more explicitly and note that the classifier's performance on other topics is inferred only from the hashtag-based automated evaluation.
[Sec. 3.2.2, Step 3] The text says the cleaning step removes "exact matches to our topic of interest" (e.g., "#mentalhealth"); this is reasonable, but it would be helpful to specify whether the same cleaning is applied to the audio transcription and how the topic-of-interest list is passed to this step in the implementation.
[Sec. 4.2] The paper does not report the actual training-phase lengths for each experiment (i.e., how many TikToks were processed before the "like at least 4 of the last 10" stopping condition was met); reporting these values would help the reader understand the variability in convergence across topics.
[Sec. 6, Future Directions] The paper states that software and datasets are planned for release, but no link or repository is provided; making these available would substantially improve reproducibility and is particularly important for an auditing tool.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the claimed steering effect is an empirical, feed-level comparison using a fixed pretrained classifier, not a result forced by the paper's own definitions or fitted parameters.

full rationale

The paper's central claim is empirical rather than definitional: a policy that likes videos when a fixed, externally pretrained zero-shot classifier exceeds a threshold is shown to produce a feed with more videos that the same classifier scores above threshold, relative to a skip-only control. The outcome is measured on the recommended feed, not on the agent's own actions, and the control is labeled with the same classifier; the treatment/control difference therefore depends on TikTok's recommender actually responding to likes and is not true by construction. The classifier (BART-large-MNLI with Whisper transcription) was not fitted to the live data, thresholds were chosen from the separate offline AUTOLIKE-D1 validation, and no fitted parameter is later renamed as a prediction. The shared-classifier design does create a construct-validity limitation: the live-feed labels are not independently verified, so the interpretation of the measured increase as an increase in genuinely harmful content is less certain. That is a correctness or external-validity concern, not circularity, because the steering result is contingent on TikTok's behavior and could in principle have failed. No load-bearing self-citations, imported uniqueness theorems, or ansatz-smuggling via citations are present. The derivation chain is therefore self-contained with respect to circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The framework rests on hand-set thresholds and horizons, on the accuracy of two off-the-shelf models (Whisper and BART-large-MNLI), and on the assumption that refreshing identifiers produces a new-user baseline. No new physical or formal entities are introduced.

free parameters (4)

Classification confidence threshold for liking = 0.5 (single-dimension), 0.6 (multi-dimension)
Chosen by hand in Sec 4.2 to decide which TikToks to like and which count as on-topic; the multi-dimension experiments raise the threshold to 0.6, and results depend on this cutoff.
Stopping condition (end goal) = like at least 4 of the last 10 TikToks
Hand-defined success criterion in Sec 4.2; it determines when the training phase ends and therefore the length and composition of each pathway.
Time horizon = 100 training plus 50 testing steps; 200 for the control
Hand-set in Sec 4.2; controls how many interactions each run uses and affects whether the recommendation system has time to personalize.
State discretization step = 0.1 grid over [0,1] x [0,1]
Sec 3.2.2 step 5 rounds classification confidence scores to the nearest valid state; the 0.1 resolution is an arbitrary modeling choice.

assumptions (5)

domain assumption The TikTok recommendation system can be treated as a Markov decision process whose relevant state is fully captured by the (topic, sentiment) classification of the current video.
Sec 2.2 defines state s as just <topic, sentiment> and uses Q-learning; this discards user history, video features, and other context that likely influence recommendations.
domain assumption A fresh For You feed, obtained by refreshing TikTok settings and resetting advertising, Android, and device IDs, represents a new-user baseline so the controlled experiment is a valid counterfactual.
Sec 3.2.1 describes the refresh procedure; the validity of the 2x comparisons in Sec 4.2 depends on this reset producing an unpersonalized feed.
domain assumption The off-the-shelf zero-shot models (BART-large-MNLI and Whisper) provide sufficiently accurate topic and sentiment labels for TikTok content in the live experiments.
Sec 3.2.2 relies on these models; manual verification in Sec 4.1 covers only a mental-health subset of the offline dataset (D1), not the live feeds (D2).
domain assumption TikTok's recommendation behavior is stationary within each experiment and is not interrupted by bot detection or account overuse flags.
Sec 4.2 notes TikTok sometimes blocks overuse or triggers bot detection; such interruptions would break the assumption that observed pathways reflect normal recommendation behavior.
standard math Q-learning on the coarse 11 by 11 discretized state space converges to a useful policy, as assumed when Eq (2) is introduced.
Sec 2.2 invokes Sutton and Barto for temporal-difference learning; no convergence argument is given for this specific state and action setup, and the deployed experiments do not run Q-learning.

how reviews work

0 comments

Cite this review

Pith. "Pith review of AutoLike: Auditing Social Media Recommendations through User Interactions." pith.science (2026). https://pith.science/paper/5CQULPAV

@misc{pith2026250208933,
  author       = {Pith},
  title        = {Pith review of: AutoLike: Auditing Social Media Recommendations through User Interactions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5CQULPAV}},
  note         = {Machine review of arXiv:2502.08933}
}

read the original abstract

Modern social media platforms, such as TikTok, Facebook, and YouTube, rely on recommendation systems to personalize content for users based on user interactions with endless streams of content, such as "For You" pages. However, these complex algorithms can inadvertently deliver problematic content related to self-harm, mental health, and eating disorders. We introduce AutoLike, a framework to audit recommendation systems in social media platforms for topics of interest and their sentiments. To automate the process, we formulate the problem as a reinforcement learning problem. AutoLike drives the recommendation system to serve a particular type of content through interactions (e.g., liking). We apply the AutoLike framework to the TikTok platform as a case study. We evaluate how well AutoLike identifies TikTok content automatically across nine topics of interest; and conduct eight experiments to demonstrate how well it drives TikTok's recommendation system towards particular topics and sentiments. AutoLike has the potential to assist regulators in auditing recommendation systems for problematic content. (Warning: This paper contains qualitative examples that may be viewed as offensive or harmful.)

Figures

Figures reproduced from arXiv: 2502.08933 by the authors.

**Figure 1.** AutoLike Framework. Formulated as a reinforcement learning problem, AutoLike enables auditing of social media recommendation algorithms. Specifically, the user provides a topic of interest and a sentiment, which characterizes the content they want to audit. A RL agent interacts with the environment (e.g., TikTok’s “For You” page). At each time step, the agent follows a RL policy to select which action to apply to th… view at source ↗

**Figure 2.** (a) We provide a conceptual example of how AutoLike can drive the RS to an end goal (red box) across two dimensions: a topic of interest (eating disorder) and sentiment (negative/sad), using real TikTok content. The RS starts at a benign state for a new user, serving popular common content, such as cats.It skips this content until it reaches more on-topic content, such as ones about dieting. During this time, it b… view at source ↗

**Figure 3.** AutoLike Implementation. We implement AutoLike for the TikTok Android app. It works in the following ways. Once the user gives the inputs, AutoLike (1) opens TikTok to its “For You” page and extracts the TikTok URL of the current recommended content; (2) sends it to a backend server and the server visits the TikTok URL in a web browser; (3) extracts the text description and transcribes the video’s audio into text, a… view at source ↗

Figures from the paper (2 more)

**Figure 5.** Figure 5: AutoLike for Single Dimensions: TikTok. We demonstrate how AutoLike can drive TikTok’s RS across individual dimensions, either for topic of interest only or for sentiment only. The y-axis denotes the number of times we liked the TikTok video (i.e., ones we classify as …

**Figure 6.** Figure 6: (a). As mental health is a broad topic, we discover a diverse range of issues displayed in the TikToks, such as body image, loss and death, depression, and self-harm [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

When Kids Mode Isn't For Kids: Investigating TikTok's "Under 13 Experience"
cs.HC 2025-06 conditional novelty 6.0 of 10

An audit of TikTok's Under 13 Experience found that 83% of For You Page videos are not child-directed, with some inappropriate content and missing parental controls.

Reference graph

Works this paper leans on

32 extracted references · 23 canonical work pages · cited by 1 Pith paper

[1]

Homeland Security & Governmental Affairs. 2021. Social Media Platforms and the Amplification of Domestic Extremism & Other Harmful Content. https://www.hsgac.senate.gov/hearings/social-media-platforms-and-the-amplification-of-domestic-extremism-and-other-harmful-content/. (Accessed on 07/24/2023)

work page 2021
[2]

Android. 2023. Write automated tests with UI Automator. https://developer.android.com/training/testing/other-components/ui-automator. (Accessed on 07/30/2023)

work page 2023
[3]

Jack Bandy and Nicholas Diakopoulos. 2023. Facebook’s News Feed Algorithm and the 2020 US Election. Social Media + Society 9 (09 2023). doi:10.1177/20563051231196898

work page doi:10.1177/20563051231196898 2023
[5]

Ashley Belanger. 2022. Parents sue TikTok after 7 kids die from profitable Blackout Challenge videos. https://arstechnica.com/tech- policy/2022/07/parents-sue-tiktok-after-7-kids-die-from-profitable-blackout-challenge-videos. (Accessed on 01/10/2025)

work page 2022
[6]

Maximilian Boeker and Aleksandra Urman. 2022. An Empirical Investigation of Personalization Factors on TikTok. InProceedings of the ACM Web Conference 2022 (Virtual Event, Lyon, France)(WWW ’22). ACM, New York, NY, USA, 2298–2309. doi:10.1145/3485447.3512102

arXiv 2022
[7]

Energy & Commerce Committee. 2023. TikTok: How Congress Can Safeguard American Data Privacy and Protect Children from Online Harms. https://energycommerce.house.gov/events/full-committee-hearing-tik-tok-how-congress-can-safeguard-american-data-privacy-and-protect- children-from-online-harms. (Accessed on 07/24/2023)

work page 2023
[8]

dfreelon. 2023. A simple module to collect video, text, and metadata from Tiktok. https://github.com/dfreelon/pyktok. (Accessed on 07/30/2023)

work page 2023
[9]

Hugging Face. 2023. facebook/bart-large-mnli. https://huggingface.co/facebook/bart-large-mnli. (Accessed on 07/30/2023)

work page 2023

Show all 32 references

[10]

Center for Countering Digital Hate. 2022. Deadly by Design. https://counterhate.com/research/deadly-by-design/. (Accessed on 07/24/2023)

2022
[11]

Daniel Funke. 2021. Why COVID-19 vaccine misinformation persists on Facebook. https://www.politifact.com/article/2021/mar/30/why-covid- 19-vaccine-misinformation-still-facebook/. (Accessed on 01/10/2025)

2021
[12]

Anna Gordon. 2024. Here’s All the Countries With TikTok Bans as Platform’s Future in U.S. Hangs In Balance. https://time.com/6971009/tiktok- banned-restrictions-worldwide-countries-united-states-law. Time (April 2024)

2024
[13]

Karen Hao. 2021. The Facebook whistleblower says its algorithms are dangerous. Here’s why. https://www.technologyreview.com/2021/10/05/ 1036519/facebook-whistleblower-frances-haugen-algorithms/?utm_source=chatgpt.com. MIT Technology Review (October 2021)

2021
[14]

Muhammad Haroon, Anshuman Chhabra, Xin Liu, Prasant Mohapatra, Zubair Shafiq, and Magdalena Wojcieszak. 2022. YouTube, The Great Radicalizer? Auditing and Mitigating Ideological Biases in YouTube Recommendations. arXiv:2203.10666 [cs.CY] 16 Le et al

2022 arXiv
[15]

Muhammad Haroon, Magdalena Wojcieszak, Anshuman Chhabra, Xin Liu, Prasant Mohapatra, and Zubair Shafiq. 2023. Auditing YouTube’s recommendation system for ideologically congenial, extreme, and problematic recommendations.Proceedings of the National Academy of Sciences 120, 50 ...

2023 doi
[16]

Rothschild, and Duncan J

Homa Hosseinmardi, Amir Ghasemian, Aaron Clauset, Markus Mobius, David M. Rothschild, and Duncan J. Watts. 2021. Examining the consumption of radical content on YouTube.Proceedings of the National Academy of Sciences 118, 32 (2021), e2101967118. doi:10.1073/pnas.2101967118

2021 doi
[17]

Dubois, David Choffnes, Athina Markopoulou, Franziska Roesner, and Zubair Shafiq

Umar Iqbal, Pouneh Nikkhah Bahrami, Rahmadi Trimananda, Hao Cui, Alexander Gamero-Garrido, Daniel J. Dubois, David Choffnes, Athina Markopoulou, Franziska Roesner, and Zubair Shafiq. 2023. Tracking, Profiling, and Ad Targeting in the Alexa Echo Smart Speaker Ecosystem. In Proc...

2023
[18]

Joel Kaplan. 2025. More Speech and Fewer Mistakes. https://about.fb.com/news/2025/01/meta-more-speech-fewer-mistakes. (Accessed on 01/09/2025)

2025
[19]

Levi Kaplan and Piotr Sapiezynski. 2024. Comprehensively Auditing the TikTok Mobile App. InCompanion Proceedings of the ACM Web Conference 2024 (Singapore, Singapore) (WWW ’24). Association for Computing Machinery, New York, NY, USA, 1198–1201. doi:10.1145/3589335.3651260

2024
[20]

Eric Kayne. 2024. Appeals Court Raises Questions Over Section 230 Law Giving Social-Media Companies Legal Immunity. https://www.wsj.com/us-ne ws/law/appeals-court-raises-questions-over-section-230-law-giving-social-media-companies-legal-immunity-af4c1e6c. (Accessed on 01/10/2025)

2024
[21]

killing people

Lilia Luciano and Grace Segers. 2021. Biden accuses social media platforms of "killing people" with spread of COVID misinformation. https://www.cbsnews.com/news/biden-facebook-social-media-covid-19-killing-people. (Accessed on 01/09/2025)

2021
[23]

I Searched for a Religious Song in Amharic and Got Sexual Content Instead

Hellina Hailu Nigatu and Inioluwa Deborah Raji. 2024. “I Searched for a Religious Song in Amharic and Got Sexual Content Instead”: Investigating Online Harm in Low-Resourced Languages on YouTube.. InProceedings of the 2024 ACM Conference on Fairness, Accountability, and Transp...

2024
[24]

OpenAI. 2023. Robust Speech Recognition via Large-Scale Weak Supervision. https://github.com/openai/whisper. (Accessed on 07/30/2023)

2023
[25]

openatx. 2023. Android Uiautomator2 Python Wrapper. https://github.com/openatx/uiautomator2. (Accessed on 07/30/2023)

2023
[26]

Brooke Perreault, Johanna Hoonsun Lee, Ropafadzo Shava, and Eni Mustafaraj. 2024. Algorithmic Misjudgement in Google Search Results: Evidence from Auditing the US Online Electoral Information Environment. In Proceedings of the 2024 ACM Conference on Fairness, Accountability, a...

2024
[27]

Jacopo Pruccoli, Marta De Rosa, Lucia Chiasso, Annalisa Perrone, and Antonia Parmeggiani. 2022. The use of TikTok among children and adolescents with Eating Disorders: experience in a third-level public Italian center during the SARS-CoV-2 pandemic.Italian Journal of Pediatric...

2022 doi
[28]

Manoel Horta Ribeiro, Raphael Ottoni, Robert West, Virgílio A. F. Almeida, and Wagner Meira. 2020. Auditing Radicalization Pathways on YouTube. In Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (Barcelona, Spain) (FAT* ’20). ACM, New York, NY,...

2020
[29]

Ronald Robertson, Jon Green, Damian Ruck, Katherine Ognyanova, Christo Wilson, and David Lazer. 2023. Users choose to engage with more partisan news than they are exposed to on Google Search.Nature 618 (05 2023), 1–7. doi:10.1038/s41586-023-06078-5

2023 doi
[30]

Congressional Research Service. 2024. Section 230: An Overview. https://crsreports.congress.gov/product/pdf/R/R46751. (Accessed on 01/10/2025)

2024
[31]

Richard Sutton and Andrew Barto. 2018. Reinforcement learning: an introduction. The MIT Press, Cambridge, Massachusetts London, England

2018
[32]

TikTok. 2023. Introducing a way to refresh your For You feed on TikTok. https://newsroom.tiktok.com/en-us/introducing-a-way-to-refresh- your-for-you-feed-on-tiktok-us. (Accessed on 07/30/2023)

2023
[33]

Karan Vombatkere, Sepehr Mousavi, Savvas Zannettou, Franziska Roesner, and Krishna P. Gummadi. 2024. TikTok and the Art of Personalization: Investigating Exploration and Exploitation on Social Media Feeds. InProceedings of the ACM Web Conference 2024(Singapore, Singapore) (WWW...

2024
[34]

Gummadi, Elissa M

Savvas Zannettou, Olivia-Nemes Nemeth, Oshrat Ayalon, Angelica Goetzen, Krishna P. Gummadi, Elissa M. Redmiles, and Franziska Roesner. 2023. Leveraging Rights of Data Subjects for Social Media Analysis: Studying TikTok via Data Donations. arXiv:2301.04945 [cs.SI] A Ethical Con...

2023 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Homeland Security & Governmental Affairs. 2021. Social Media Platforms and the Amplification of Domestic Extremism & Other Harmful Content. https://www.hsgac.senate.gov/hearings/social-media-platforms-and-the-amplification-of-domestic-extremism-and-other-harmful-content/. (Accessed on 07/24/2023)

work page 2021

[2] [2]

Android. 2023. Write automated tests with UI Automator. https://developer.android.com/training/testing/other-components/ui-automator. (Accessed on 07/30/2023)

work page 2023

[3] [3]

Jack Bandy and Nicholas Diakopoulos. 2023. Facebook’s News Feed Algorithm and the 2020 US Election. Social Media + Society 9 (09 2023). doi:10.1177/20563051231196898

work page doi:10.1177/20563051231196898 2023

[4] [5]

Ashley Belanger. 2022. Parents sue TikTok after 7 kids die from profitable Blackout Challenge videos. https://arstechnica.com/tech- policy/2022/07/parents-sue-tiktok-after-7-kids-die-from-profitable-blackout-challenge-videos. (Accessed on 01/10/2025)

work page 2022

[5] [6]

Maximilian Boeker and Aleksandra Urman. 2022. An Empirical Investigation of Personalization Factors on TikTok. InProceedings of the ACM Web Conference 2022 (Virtual Event, Lyon, France)(WWW ’22). ACM, New York, NY, USA, 2298–2309. doi:10.1145/3485447.3512102

arXiv 2022

[6] [7]

Energy & Commerce Committee. 2023. TikTok: How Congress Can Safeguard American Data Privacy and Protect Children from Online Harms. https://energycommerce.house.gov/events/full-committee-hearing-tik-tok-how-congress-can-safeguard-american-data-privacy-and-protect- children-from-online-harms. (Accessed on 07/24/2023)

work page 2023

[7] [8]

dfreelon. 2023. A simple module to collect video, text, and metadata from Tiktok. https://github.com/dfreelon/pyktok. (Accessed on 07/30/2023)

work page 2023

[8] [9]

Hugging Face. 2023. facebook/bart-large-mnli. https://huggingface.co/facebook/bart-large-mnli. (Accessed on 07/30/2023)

work page 2023