Crowdsourced audit of Twitter’s recommender systems
Open accessAbstract
This research conducts an audit of Twitter’s recommender system, aiming to examine the disparities between users’ curated timelines and their subscription choices. Through the combined use of a browser extension and data collection via the Twitter API, our investigation reveals a high amplification of friends from the same community, a preference for amplifying emotionally charged and toxic tweets and an uneven algorithmic amplification across friends’ political leaning. This audit emphasizes the importance of transparency, and increased awareness regarding the impact of algorithmic curation.
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
Approximately 50% of tweets in X's user timelines are personalized recommendations from accounts users do not follow.
“Approximately 50% of tweets in X's user timelines are personalized recommendations from accounts they do not follow.”
From Auditing Political Exposure Bias: Algorithmic Amplification on Twitter/X During the 2024 U.S. Presidential Election (2025) · cites this paper as [8]
The Methods section (‘Quantification of algorithmic amplification’; page not available) confirms that ‘For You’ timelines include tweets from accounts users do not follow. But it reports that these tweets made up approximately 37% of timelines in the study cohort, not approximately 50%. (LLM, conf=0.88)
Checked by a reasoning model — may be wrong.
Experimental — accuracy not yet measuredEarlier research suggests that personalized recommendations tend to be more centrist in political stance.
“However, our findings diverge from earlier research suggesting that personalized recommendations tend to be more centrist in political stance [8,9,29].”
From Auditing Political Exposure Bias: Algorithmic Amplification on Twitter/X During the 2024 U.S. Presidential Election (2025) · cites this paper as [8]
The Results section (‘Algorithmic curation affects the political landscape’; page unavailable) reports that Twitter amplified ideologically aligned accounts for far-left and left/center-left participants, while center-right participants saw greater amplification of far-left and further-right accounts. These findings do not describe a tendency toward more centrist recommendations; the Discussion instead emphasizes amplification of content consonant with users’ communities. (LLM, conf=0.85)
Checked by a reasoning model — may be wrong.
Experimental — accuracy not yet measuredOther studies report increased exposure to ideologically aligned friends but decreased exposure to external links.
“Other studies report increased exposure to ideologically aligned friends [4,8], but decreased exposure to external links [4,29].”
From Auditing Political Exposure Bias: Algorithmic Amplification on Twitter/X During the 2024 U.S. Presidential Election (2025) · cites this paper as [8]
The Results section, “Algorithmic curation affects the political landscape,” reports amplification of ideologically aligned friends for far-left and left/center-left participants, but the opposite pattern for center-right participants. Our copy does not report decreased exposure to external links; searches for external links, URLs, and hyperlinks found no supporting passage. No page numbers were available in the indexed copy. (LLM, conf=0.85)
Experimental — accuracy not yet measuredBouchaud et al. (2023) found that users viewed an average of 30 tweets in one session of using Twitter.
“Looking at the first ten tweets might be a reasonable approximation though-Bouchaud et al. (2023) found that users viewed an average of 30 tweets in one session of using Twitter [9].”
From Twitter's Algorithm: Amplifying Anger, Animosity, and Affective Polarization (2023) · cites this paper as [9]
Methods > Quantification of algorithmic amplification discusses tweets displayed during a session, but says the authors aggregated displayed tweets over two weeks instead. Neither that passage nor the searched Data collection passages report an average of 30 tweets viewed per session (searched: “30 tweets,” “average per session,” and “tweets seen per browsing session”). Page numbers were not available in our copy. (LLM, conf=0.72)
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Cites
Exposure to opposing views on social media can increase political polarization
Bail, C. A., Argyle, L. P., Brown, T. W., Bumpus, J. P., Chen, H., Hunzaker, M. B. F., Lee, J., Mann, M., Merhout, F., & Volfovsky, A. (2018) · Proceedings of the National Academy of Sciences · doi:10.1073/pnas.1804840115 · Cited by 1,863 · see citing works
Flagged first, then by citations
Research works printed but not linked: 19
- [1]Ovadya, A. & Thorburn, L. Bridging systems: Open problems for countering destructive divisiveness across ranking, recommenders, and governance (2023). arxiv: 2301.09976. · arXiv
- [2]Huszár F Algorithmic amplification of politics on twitter Proc. Natl. Acad. Sci. U.S.A. 2021 10.1073/pnas.2025334119 34934011 PMC8740571 · DOI · Paper page
- [3]Bandy J Diakopoulos N More accounts, fewer links Proc. ACM Hum. Comput. Interact. 2021 5 1 28 10.1145/3449152 · DOI · Paper page
- [4]Bartley, N., Abeliuk, A., Ferrara, E. & Lerman, K. Auditing algorithmic bias on twitter. In 13th ACM Web Science Conference 2021, 10.1145/3447535.3462491 (ACM, 2021). · DOI · Paper page
- [5]Hargreaves, E. et al. Biases in the facebook news feed: A case study on the italian elections. In 2018 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), 10.1109/asonam.2018.8508659 (IEEE, 2018). · DOI
- [6]Sanna, L., Romano, S., Corona, G. & Agosti, C. YTTREX: Crowdsourced analysis of YouTube’s recommender system during COVID-19 pandemic. In Information Management and Big Data, 107–121, 10.1007/978-3-030-76228-5_8 (Springer International Publishing, 2021). · DOI
- [7]Milli, S. et al. Engagement, user satisfaction, and the amplification of divisive content on social media (2023). arxiv: 2305.16941. PMC11894805 40070432 · arXiv
- [8]Wagner C Measuring algorithmically infused societies Nature 2021 595 197 204 10.1038/s41586-021-03666-1 34194046 · DOI · Paper page
- [10]Satuluri, V. et al. Simclusters. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, 10.1145/3394486.3403370 (ACM, 2020). · DOI
- [11]Grover, A. & Leskovec, J. node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 10.1145/2939672.2939754 (ACM, 2016). PMC5108654 27853626 · DOI · Paper page
- [12]McInnes L Healy J Astels S hdbscan: Hierarchical density based clustering J. Open Source Softw. 2017 10.21105/joss.00205 · DOI · Paper page
- [14]Davidson, T., Bhattacharya, D. & Weber, I. Racial bias in hate speech and abusive language detection datasets (2019). arxiv: 1905.12516. · arXiv
- [16]Gaumont N Panahi M Chavalarias D Reconstruction of the socio-semantic dynamics of political activist twitter networks-method and application to the 2017 french presidential election PLoS One 2018 13 e0201879 10.1371/journal.pone.0201879 30231018 PMC6145593 · DOI · Paper page
- [17]Chavalarias, D., Bouchaud, P. & Panahi, M. Can few lines of code change society? beyond fack-checking and moderation: how recommender systems toxifies social networking sites. (under review)10.48550/arXiv.2303.15035 (2023). · DOI
- [18]Kojaku, S., Radicchi, F., Ahn, Y.-Y. & Fortunato, S. Network community detection via neural embeddings (2023). arxiv: 2306.13400. PMC11530665 39487114 · arXiv
- [19]Jolly S Chapel hill expert survey trend file, 1999–2019 Elect. Stud. 2022 75 102420 10.1016/j.electstud.2021.102420 · DOI
- [20]Barbieri, F., Espinosa Anke, L. & Camacho-Collados, J. XLM-T: Multilingual language models in Twitter for sentiment analysis and beyond. In Proceedings of the Thirteenth Language Resources and Evaluation Conference, 258–266 (European Language Resources Association, Marseille, France, 2022).
- [22]Lu Y Myrick JG Cross-cutting exposure on Facebook and political participation J. Media Psychol. 2016 28 100 110 10.1027/1864-1105/a000203 · DOI
- [24]Schneider FM Weinmann C In need of the devil’s advocate? The impact of cross-cutting exposure on political discussion Polit. Behav. 2021 45 373 394 10.1007/s11109-021-09706-w · DOI
Cited by
Sorted by citations
Pedreschi, D., Pappalardo, L., Ferragina, E., Baeza-Yates, R., Barabási, A.-L., Dignum, F., Dignum, V., Eliassi-Rad, T., Giannotti, F., Kertész, J., et al. (2025) · Artificial Intelligence · doi:10.1016/j.artint.2024.104244 · Cited by 102 · see citing works
Milli, S., Carroll, M., Wang, Y., Pandey, S., & Dragan, A. D. (2023) · arXiv.org · doi:10.48550/arxiv.2305.16941 · Cited by 200 · see citing works
Ye, J., Luceri, L., & Ferrara, E. (2025) · Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency · doi:10.1145/3715275.3732159 · Cited by 16 · see citing works
“As of this writing, X ’s “For You” timeline typically consists of 50% in-network tweets (i.e., from accounts a given user follows) and 50% out-of-network tweets (i.e., from accounts that user does not directly follow)—an increase from the 37% out-of-network tweets in 2023 [8].”
Influential
- Authors
- Bouchaud, P., Chavalarias, D., & Panahi, M.
- Affiliations
- Centre National de la Recherche Scientifique (FR)
- Corresponding author
- Bouchaud, Paul
- Venue
- Scientific Reports
- Published
- 2023-10-05
- Volume
- 13
- Issue
- 1
- Pages
- 16815-16815
- ISSN
- 2045-2322 (ISSN-L)
- PMC
- PMC10556069
- OpenAlex
- W4387277768
- Cited by
- 21
- References
- 1
- Citations by year
- 2026: 6 · 2025: 8 · 2024: 7
- FWCI
- 5.25
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Locations
- Scientific Reports (publishedVersion, cc-by)
- PubMed Central (publishedVersion, cc-by)
- PubMed (publishedVersion)
- HAL (submittedVersion, other-oa)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Funders
- Fondation CFM pour la recherche
- Region Île-de-France
- Topics
- Social Sciences › Safety Research › Ethics and Social Impacts of AI
- Social Sciences › Sociology and Political Science › Misinformation and Its Impacts
- Social Sciences › Communication › Social Media and Politics
- Keywords
- Twitter, recommender systems, algorithmic amplification, social media curation, algorithmic transparency, toxic content, political polarization
- MeSH terms
- Communication, Humans, Social Media, Crowdsourcing
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
