CoChat
Sign Up
Cited by 16Social Media and PoliticsEthics and Social Impacts of AIMisinformation and Its Impacts

Auditing Political Exposure Bias: Algorithmic Amplification on Twitter/X During the 2024 U.S. Presidential Election

Open access
Ye, Jinyi, Luceri, Luca, Ferrara, Emilio·2025·Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency
DOI 10.1145/3715275.3732159
Open this paper in CoChat — free accountOpen PDF

Abstract

Approximately 50% of tweets in X's user timelines are personalized recommendations from accounts they do not follow.This raises a critical question: What political content are users exposed to beyond their established networks, and what implications does this have for democratic discourse online?In this paper, we present a six-week audit of X's algorithmic content recommendations during the 2024 U.S. Presidential Election by deploying 120 sock-puppet monitoring accounts to capture tweets from their personalized "For You" timelines.Our objective is to quantify out-of-network content exposure for right-and left-leaning account profiles and assess any potential inequalities and biases in political exposure.Our findings indicate that X's algorithm skews exposure toward a few highpopularity users across all monitoring accounts, with right-leaning accounts experiencing the highest level of exposure inequality.Both left-and right-leaning accounts encounter amplified exposure to users aligned with their own political views and reduced exposure to opposing viewpoints.Additionally, we observe that new accounts experience a right-leaning bias in exposure within their default timelines.Our work contributes to understanding how content recommendation systems may induce and reinforce biases while exacerbating vulnerabilities among politically polarized user groups.We underscore the importance of transparency-aware algorithms in addressing critical issues such as safeguarding election integrity and fostering a more informed digital public sphere.

AI snapshot

Written from the paper’s full text.

TL;DR

A six-week sock-puppet audit of X’s “For You” timeline during the 2024 U.S. Presidential Election finds that out-of-network recommendations are concentrated among a few high-popularity users and are ideologically aligned with partisan accounts. Neutral/new accounts show a default right-leaning exposure bias.

Key findings

  • Exposure inequality is moderate-to-high across all groups: average Gini coefficient exceeds 0.45, with right-leaning accounts experiencing the highest inequality, followed by balanced and left-leaning accounts; all pairwise differences are significant at p < 0.001.
  • Neutral accounts’ top recommendations skew right: among the top 20 recommended users, right-leaning users make up 30.16% of exposure versus 12.92% for left-leaning users; this persists in the top 50 (35.26% vs. 22.34%) and top 100 (31.39% vs. 20.83%).
  • Partisan accounts see ideologically aligned amplification and opposing-viewpoint de-amplification: top liberal and conservative voices are amplified more than 50% above the balanced-account baseline for left- and right-leaning accounts, respectively. Mean amplification is 36.76% for left-leaning accounts and 30.29% for right-leaning accounts (Mann-Whitney U p < 0.05), with no significant difference in de-amplification between groups.
  • Out-of-network content is a large share of timelines: neutral accounts see 100% out-of-network tweets, while left-, right-, and balanced accounts see 55%–63% out-of-network tweets.

Methods

  • Deployed 120 sock-puppet X accounts in four groups of 30: neutral (following no one), left-leaning, right-leaning, and balanced. Political orientation was set via follows based on the AllSides Media Bias Chart; left/right accounts followed 10 media outlets plus Democratic or Republican figures, while balanced accounts followed center-left and center-right media and both presidential candidates.
  • A timeline crawler collected “For You” timelines four times daily from October 2 to November 19, 2024, yielding 9.79 million tweets; each session captured roughly 500–700 tweets, or about 2,000–3,000 tweets per account per day.
  • Analyses used a “weighted occurrence per 1,000 tweets” exposure metric with exponential rank decay, Gini coefficients and Lorenz curves for inequality, and a mean amplification ratio comparing partisan accounts to the balanced-account baseline, with Mann-Whitney U tests.

Limitations

  • The paper states the six-week election-period window may limit reproducibility in less politically sensitive periods; the no-interaction sock-puppet design does not capture personalization or user-algorithm interaction effects; pre-selected interests, age, and location may still confound neutral-account recommendations; and the balanced-account baseline may not generalize across diverse demographics or global political contexts.
  • The text also notes that political leanings of recommended users were inferred from publicly available data and “may be subject to inaccuracies or changes over time.”
  • The paper acknowledges ecological-validity concerns about sock-puppet audits but argues controlled accounts are appropriate for auditing baseline algorithmic recommendations.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1145/3715275.3732159

Cites

31 of 31 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

18 of 29 cited sources checked · 8 supported · 2 flagged · accuracy not yet measured

  • Premisecitation

    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.”
    • Also stated in: § 1 Introduction · plain
      “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].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [8] source

      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 measured
  • Interpretationcitation

    Algorithmic recommendations can reinforce echo chambers not just in the composition of social networks but also in the ideological framing of content circulating in the network.

    “The X algorithm's amplification of ideologically aligned out-of-network accounts, along with the reduced exposure to opposing viewpoints, suggests that algorithmic recommendations can reinforce echo chambers not just in the composition of social networks [12] but also in the ideological framing of content circulating in the network.”
    Quote in the paperSubstantial claim1 of 1 sources checked
  • Interpretationcitation

    The increased prominence of non-institutional voices such as verified political commentators and influencers potentially introduces sensationalism and misinformation into echo chambers.

    “The increased prominence of non-institutional voices, such as verified political commentators and influencers, further exacerbates this issue by potentially introducing sensationalism and misinformation into these echo chambers [11].”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites [11] source

      The Introduction says verification can increase users’ visibility and reach, and the Discussion links priority users to stronger echo chambers. But the paper does not identify non-institutional political commentators or influencers as the relevant users, or show that they introduce sensationalism or misinformation. The Introduction explicitly says misinformation falls outside the study’s scope; no page numbers are available in our copy. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
  • Premisecitation

    Algorithms on social media platforms can amplify selective exposure by recommending content similar to what users already prefer or agree with.

    “Algorithms on social media platforms can amplify this effect by recommending content similar to what users already prefer or agree with, reinforcing selective exposure through personalization [22].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [22] source

      The section on NRSs’ causal effects says recommenders can promote articles resembling those users previously chose and describes a possible feedback loop that could increase selective exposure. But the study focuses on online news sites; the Discussion says its framework may be relevant to social media platforms, not that it establishes the claimed effect there. Page numbers are not available in our copy. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
  • Premisecitation

    Previous sock-puppet audits on X are often constrained by small sample sizes of fewer than 10 accounts and restricted tweet collection in quantity and frequency, limiting the generalization and robustness of their findings.

    “Furthermore, previous sock-puppet audits on X are often constrained by small sample sizes (fewer than 10 accounts) and restricted tweet collection in terms of both quantity and frequency [4,6,9], limiting the generalization and robustness of their findings.”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [4] source
    • cites [6] source

      Section 3.2 (Audit Bots; page unavailable) reports that this audit used eight accounts because of scaling issues and collected at least 15 tweets per session, six times a day. Section 3.3 (Timeline Scraping; page unavailable) says the observed tweets represented less than 10% of content produced by the accounts the bots followed. These passages support the description of this audit’s scale and collection limits, but do not establish that previous audits were often similarly constrained or state that those limits reduced the generalizability and robustness of their findings. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [9]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Classified as probably not a substantial claim. 19
  • May combine two claims. 25

  • Classified as probably not a substantial claim.

    Premisecitation

    Earlier 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].”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites [8] source

      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 measured
    • cites [9]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [29] source

  • May combine two claims.

    Interpretationcitation

    Even beyond the friend network, exposure inequality remains at a similar level to that reported for friends' tweets, indicating that the platform's algorithm amplifies certain accounts both within and outside of users' direct networks.

    “This suggests that even beyond the friend network, exposure inequality remains at a similar level, indicating that the platform's algorithm amplifies certain accounts both within and outside of users' direct networks.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [6] source

      Section 4.3.1 reports that exposure to friends’ tweets is more unequal in the personalized timeline, and Section 5 attributes a disproportionate share of that timeline to a few friends. Neither section reports exposure inequality beyond the friend network or compares its level with inequality for friends’ tweets. Page numbers are unavailable in our copy. (LLM, conf=0.87)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Previous research has shown that popularity biases can lead to a skew in the visibility of tweets when comparing personalized feeds with reverse-chronological ones, and that users are disproportionately exposed to friends' tweets.

    “Previous research has shown that popularity biases can lead to a skew in the visibility of tweets when comparing personalized feeds with reverse-chronological ones, and that users are disproportionately exposed to friends' tweets [6,7].”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [6] source
    • cites [7] source

      RQ1 reports a skew in the popularity of tweets observed in personalized versus reverse-chronological feeds. The Introduction describes how highly active followed accounts can be overrepresented, but RQ1 says the eight-sockpuppet study did not identify a skew in exposure between the activity network and the feeds. The paper therefore does not establish the broader claim that users are disproportionately exposed to friends’ tweets. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Other 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].”
    Quote in the paperNot a substantial claim2 of 3 sources checked
    • cites [4] source
    • cites [8] source

      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 measured
    • cites [29] source

  • May combine two claims.

    Interpretationcitation

    Current recommendation systems appear to disproportionately amplify high-popularity accounts, creating inequality in exposure that may result in less personalized and miscalibrated recommendations for certain user groups.

    “Current recommendation systems appear to disproportionately amplify high-popularity accounts, creating inequality in exposure that may result in less personalized and miscalibrated recommendations for certain user groups [1].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [1] source

      Section 3.2 (page not supplied) says algorithms often amplify popularity bias by overrecommending popular items, including to users interested in niche items. Section 5.1 links this bias to unequal effects and miscalibration across user groups. The paper discusses items, not high-popularity accounts, and does not establish unequal exposure for accounts. (LLM, conf=0.88)

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Previous studies on algorithmic bias have reported amplification patterns within users' in-network content similar to the exposure inequality found in this study.

    “Our analysis of exposure inequality aligns with previous studies on algorithmic bias, which have reported similar amplification patterns within users' in-network content [7].”
    Quote in the paperNot a substantial claim1 of 1 sources checked

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Some studies have found that Twitter's algorithms tend to amplify content from right-leaning media sources and politicians more than their left-leaning counterparts.

    “Some have found that Twitter's algorithms tend to amplify content from right-leaning media sources and politicians more than their left-leaning counterparts [15,20].”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [15]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [20] source

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Studies observe increased low-credibility content in algorithmic timelines, with right-leaning users experiencing higher exposure to such content.

    “Studies also observe increased low-credibility content in algorithmic timelines [10], with right-leaning users experiencing higher exposure to such content [9].”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [10] source

      Section 3.2 (page unavailable) reports greater amplification of right-leaning low-credibility tweets, which supports increased visibility of that content. The citation is attached to a claim about right-leaning users, however; the paper classifies the political bias of content, not the political leaning of users receiving it, so it does not establish that those users had higher exposure. (LLM, conf=0.86)

      Experimental — accuracy not yet measured
    • cites [9]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Bakshy et al. report considerable cross-cutting exposure on Facebook, and Wang et al. find that Twitter/X provides higher-quality and less ideologically congruent news curation.

    “Bakshy et al. [2] report considerable cross-cutting exposure on Facebook, and Wang et al. [29] find that Twitter/X provides higher-quality and less ideologically congruent news curation.”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites [2] source
    • cites [29] source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Haroon et al. found that YouTube's algorithm consistently promotes ideologically aligned content to partisan users.

    “Haroon et al. [18] trained sock puppets to represent five ideological positions ranging from left to right, and found that YouTube's algorithm consistently promotes ideologically aligned content to partisan users.”
    Quote in the paperNot a substantial claim1 of 1 sources checked

  • May combine two claims.

    Findingown data

    Neutral accounts' timelines consist exclusively of out-of-network tweets, while left-leaning, right-leaning, and balanced accounts have 55%-63% of their timelines composed of out-of-network tweets.

    “It shows the average proportion of out-of-network tweets that each account type encounters, with neutral accounts seeing exclusively out-of-network content, while the other accounts have 55%-63% of their timelines composed of out-of-network tweets.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    New accounts, or neutral accounts that follow no one, experience a right-leaning bias in exposure within their default timelines.

    “Additionally, we observe that new accounts experience a right-leaning bias in exposure within their default timelines.”
    • Also stated in: § 1 Introduction · plain
      “Additionally, neutral accounts who do not follow anybody (akin to a newly-registered user account) show a default right-leaning bias in content exposure.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Balanced accounts receive a roughly even mix of left- and right-leaning recommendations, whereas left- and right-leaning accounts predominantly receive recommendations from ideologically aligned users.

    “Notably, balanced accounts receive a roughly even mix of leftand right-leaning recommendations, whereas left-and right-leaning accounts predominantly receive recommendations from ideologically aligned users.”
    • Also stated in: § B Top Recommended Users in Left-Leaning, Right-Leaning, and Balanced Accounts · plain
      “Notably, balanced accounts receive a roughly even mix of leftand right-leaning recommendations, whereas left-and right-leaning accounts predominantly receive recommendations from ideologically aligned users.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Left-leaning sock-puppet accounts tend to see left-leaning users amplified and right-leaning users de-amplified, with the opposite pattern observed for right-leaning accounts.

    “A qualitative inspection reveals that left-leaning sock-puppet accounts tend to see left-leaning users amplified, and right-leaning users de-amplified, with the opposite pattern observed for rightleaning accounts.”
    • Also stated in: § 5 Discussion & Conclusions · plain
      “Both left-and right-leaning accounts see amplified exposure to users aligned with their political stance, while exposure to opposing viewpoints is reduced.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S1 (asserted), S3 (load-bearing), S4 (based).

    Findingnone given

    For right-leaning accounts, the most de-amplified accounts are JoJoFromJerz, acnewsitics, and Tristan Snell.

    “Conversely, the most de-amplified accounts are JoJoFromJerz (a left-leaning political influencer), acnewsitics (a liberal-leaning news commentator), and Tristan Snell (a pro-democrat lawyer and legal commentator).”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Top liberal and conservative voices are amplified more than 50% above baseline for left- and right-leaning users, respectively.

    “As shown in Figure 4, top liberal and conservative voices are amplified more than 50% above baseline for left-and right-leaning users, respectively.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Out-of-network exposure inequality differs significantly in all pairwise comparisons between account groups (Mann-Whitney U test, p < 0.001), with right-leaning users experiencing the highest out-of-network exposure inequality.

    “Significant disparities are found in all pairwise comparisons (Mann-Whitney U test: 𝑝 < 0.001), with right-leaning users experiencing the highest out-of-network exposure inequality.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    X skews exposure toward a select few high-popularity users for all accounts, with right-leaning accounts experiencing the highest level of inequality.

    “Using 120 sock-puppet accounts with left-leaning, right-leaning, balanced, and neutral political orientations, we observe that X skews exposure toward a select few high-popularity users for all accounts, with right-leaning accounts experiencing the highest level of inequality.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Twitter's ranking algorithm employs a ~48M parameter neural network that uses thousands of features to score each tweet based on engagement probabilities, prioritizing content with higher likelihoods of interaction in users' feeds.

    “Twitter's ranking algorithm employs a ~48M parameter neural network, which uses thousands of features to score each tweet based on engagement probabilities, prioritizing content with higher likelihoods of interaction in users' feeds [28].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [28]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    In 2023, Twitter partially open-sourced its recommendation algorithm, revealing that out-of-network recommendations are sourced through engagement and follow graphs, ranked by a neural network, and refined with heuristics and filters.

    “In 2023, Twitter partially open-sourced its recommendation algorithm, revealing that out-of-network recommendations are sourced through engagement and follow graphs, ranked by a neural network, and refined with heuristics and filters [28].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [28]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductown data

    The study used a program to randomly select interests and random birthdates between 1990 and 1999 for sock-puppet accounts.

    “We use a program to select these interest randomly, alongside random birthdates between 1990 and 1999.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    The algorithm tends to amplify conservative figures more heavily in right-leaning timelines while reducing exposure to left-leaning accounts.

    “This pattern highlights the algorithm's tendency to amplify conservative figures more heavily in right-leaning timelines while reducing exposure to left-leaning accounts.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Content recommendation systems can influence and amplify biases, potentially increasing vulnerabilities within politically polarized user groups.

    “Our findings reveal how content recommendation systems can influence and amplify biases, potentially increasing vulnerabilities within politically polarized user groups.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductreasoning

    A VPN was used during data collection to randomize account attributes and mitigate location-based biases in recommendations.

    “To further randomize account attributes and mitigate location-based biases in recommendations, a VPN was used during data collection.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductreasoning

    Because the sock-puppet accounts do not engage with tweets, the study does not capture the effects of user-algorithm interactions on political exposure bias and does not account for personalization or the dynamics of user activity.

    “Although our sock-puppet auditing method ensures precise control over account behaviors, it does not account for personalization or the dynamics of user activity.Since the sock-puppet accounts do not engage with tweets (e.g., clicking, responding, or retweeting), our study does not capture the effects of user-algorithm interactions on political exposure bias.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The use of balanced accounts as a baseline for measuring exposure biases may not fully capture the platform's broader algorithmic behavior across diverse user demographics or global political contexts, potentially limiting the generalizability of the findings.

    “Forth, the use of balanced accounts as a baseline for measuring exposure biases may not fully capture the platform's broader algorithmic behavior across diverse user demographics or global political contexts, potentially limiting the generalizability of our findings.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductreasoning

    Potential confounding factors such as pre-selected interests, age, and location may influence algorithmic recommendations for neutral accounts, and although these settings were randomized, their residual effects cannot be entirely ruled out.

    “Third, potential confounding factors, such as pre-selected interests, age, and location, may influence algorithmic recommendations for neutral accounts.While we took care to randomize these settings, their residual effects cannot be entirely ruled out.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductreasoning

    Because the platform disabled the timeline immediately after account creation for accounts that followed no users during the election period, data collection was limited to approximately 10 older neutral accounts after November 5.

    “However, during the election period, the platform temporarily modified this restriction, disabling the timeline immediately after account creation for accounts that followed no users.Consequently, data collection was limited to approximately 10 older neutral accounts after November 5.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Left-leaning timelines prioritize left-aligned figures while downplaying right-leaning accounts.

    “This suggests that left-leaning timelines prioritize leftaligned figures while downplaying right-leaning accounts.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductreasoning

    The study was conducted during a six-week period leading up to the 2024 U.S. elections, a politically charged time that may differ from other contexts and could affect the reproducibility of the results in less politically sensitive periods.

    “First, the study is conducted during a six-week period leading up to the 2024 U.S. elections, a politically charged time that may differ from other contexts.This temporal limitation could affect the reproducibility of our results in less politically sensitive periods.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation

    Findings from TikTok and YouTube indicate that the top 20% of an account's videos receive more than 70% of the views.

    “The parameters of the exponential decay function are informed by findings from studies on platforms like TikTok and YouTube [17], which indicate that the top 20% of an account's videos receive more than 70% of the views.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [17] source

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Study design and conductnone given

    Certain factors, such as X's default settings or trending topics, could still introduce slight biases into neutral accounts' recommendations.

    “However, it is worth noting that certain factors, such as X's default settings or trending topics, could still introduce slight biases into these recommendations.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Study design and conductreasoning

    The configuration of neutral accounts limits bias from following choices and captures a baseline view of algorithm behavior when no user preferences are specified.

    “This configuration limits any bias that could arise from following choices, aiming to capture a baseline view of how the algorithm behaves when no user preferences are specified.”
    Quote in the paperSubstantial claim
33 of 33

  • The authors state this as an assumption.

    The top 20% of tweets in a timeline capture approximately 70% of user attention.

    “Using this as a reference, we assume that the top 20% of tweets in a timeline similarly capture the majority (70%) of user attention, and we calibrate our decay model accordingly.”

  • The paper relies on this without support.

    A politically balanced set of followed accounts provides a valid baseline for measuring algorithmic amplification.

    “we introduce a self-constructed baseline using accounts that follow a politically balanced social media diet, enabling direct comparisons with other manually-created partisan user accounts.”

  • The paper relies on this without support.

    Neutral accounts that follow no one provide an unbiased look at the algorithm's default behavior.

    “Here, we are particularly interested in neutral accounts, which provide an unbiased look at the algorithm's default behavior.”

  • The paper relies on this without support.

    Sock-puppet accounts with controlled follows accurately represent real user behavior and interactions for the purpose of auditing algorithmic recommendations.

    “One general concern about sockpuppet audits is ecological validity-whether artificial accounts accurately represent real user behavior and interactions [3,4,29].”

  • The paper relies on this without support.

    The absence of user interactions captures inherent biases in the recommendation system.

    “our goal is to capture inherent biases in the recommendation system, best analyzed without user interactions.”

  • The paper relies on this without support.

    The political leaning of recommended users can be reliably inferred from publicly available data.

    “To further illustrate this trend, usernames are displayed in blue (left-leaning) or red (right-leaning) based on their political stance, which is inferred from publicly available data (may be subject to inaccuracies or changes over time).”
    “Political leanings of users are inferred based on publicly available data, which may be subject to inaccuracies or changes over time.”

  • The paper relies on this without support.

    The timelines of sock-puppet accounts that follow only a few moderately partisan media and politicians are representative of what a real new user would see.

    “Given that our sockpuppet accounts only follow a few moderately partisan media and politicians, it suggests that once a new user begins following a few partisan accounts, their algorithmic recommendations quickly become filled with like-minded voices.”
7 of 7

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • The stated six-week audit spans about seven weeks

    Minor

    October 2 through November 19 is 49 calendar days inclusive, or seven weeks, rather than six.

1 of 1

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Ye, J., Luceri, L., & Ferrara, E.
Affiliations
University of Southern California (US)
ORCID
Venue
Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency
Publisher
ACM
Published
2025-06-23
Pages
2349-2362
DOI
10.1145/3715275.3732159
OpenAlex
W4411542040
Semantic Scholar
64be8540f06a91a7cd0e177cfe7fc9d39b436219
Cited by
16
References
29
Citations by year
2026: 12 · 2025: 3 · 2024: 1
FWCI
107.38
Type
conference paper
Language
en
Open access
Licence
cc-by
Locations
  • ACM Conference on Fairness, Accountability, and Transparency (FAccT) (publishedVersion, cc-by)
  • arXiv (Cornell University) (submittedVersion, cc-by)
  • arXiv (Cornell University) (cc-by)
  • Zenodo (CERN European Organization for Nuclear Research) (cc-by)
Funders
NSF (2331722)
Topics
  • Social Sciences › Communication › Social Media and Politics
  • Social Sciences › Safety Research › Ethics and Social Impacts of AI
  • Social Sciences › Sociology and Political Science › Misinformation and Its Impacts
Keywords
algorithmic amplification, Twitter, 2024 US presidential election, personalized content recommendation, For You Page, political polarization, echo chambers, electoral integrity, algorithmic transparency

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.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-13.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-13.

Cited sources are checked by gpt-6-sol.

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