Tonneau, M., Liu, D., McGrady, R., Zheng, K., Schroeder, R., Zuckerman, E., & Hale, S. (2026). Language Disparities in Moderation Workforce Allocation by Social Media Platforms. Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency. https://doi.org/10.1145/3805689.3806484

View paper

Summary

This paper offers the first comparative empirical analysis of how six major social media platforms — YouTube, Meta, TikTok, Twitter/X, Snapchat, and LinkedIn — distribute human content moderators across languages. Drawing on transparency reports mandated by the EU’s Digital Services Act (DSA), the authors document substantial cross-lingual disparities in both language coverage (whether a language has any human moderators at all) and intensity (moderators relative to content volume). They find that smaller platforms leave millions of EU users without moderation in their national language, and that languages predominantly spoken in the Global South — Spanish, Portuguese, Arabic — receive consistently fewer moderators per unit of content than English. The paper argues that DSA transparency, while a pivotal advance, remains insufficient for independent scrutiny because platforms do not disclose content volume per language, and calls for more meaningful, globally inclusive reporting requirements.

Key Contributions

  • First comparative empirical analysis of cross-lingual moderator workforce allocation across six major DSA-regulated platforms.
  • A novel methodology combining DSA transparency disclosures with independently constructed, calibrated content-volume estimates to normalize moderator counts by language.
  • Quantification of the number of EU users left without national-language moderation on specific platforms.
  • Empirical grounding that extends prior whistleblower and NGO reports (e.g., Haugen disclosures, Global Witness) on non-English underinvestment.
  • Concrete policy recommendations: require platforms to report content volume per language, moderator workload capacity, and harmful-content prevalence, and to extend transparency obligations beyond the EU.

Methods

The authors collected moderator counts per language from DSA-mandated transparency reports (Summer 2023–Fall 2024), averaging across reporting periods and treating non-reported EU languages as having zero moderators. To normalize by content volume, they built independent language-share estimates: the TwitterDay dataset (375M tweets from one day in 2022) for Twitter/X and a random sample of ~26,000 YouTube videos for YouTube. Language identification used fastText on tweet text and video titles/descriptions, with inference scores calibrated via isotonic regression against native-speaker annotations, and bootstrap resampling (1,000 samples) for confidence intervals. Twitter/X user locations were mapped to countries via the Google Geocoding API, and DSA-reported EU user counts were used to estimate how many users lack national-language moderation, categorized with the UN geoscheme.

Findings

  • YouTube, Meta, and TikTok cover nearly all EU official languages; Twitter/X, LinkedIn, and Snapchat have multiple blind spots, especially in Southern, Eastern, and Northern Europe.
  • Roughly 16M EU Twitter/X users (14%), 8M LinkedIn users (16%), and 7M Snapchat users (7%) have no moderators for their national language.
  • On Twitter/X, blind-spot languages account for an average 31% of tweets in countries where they are official.
  • On Twitter/X, only Bulgarian and German exceed English in moderators-per-content; Portuguese, Arabic, and Spanish each receive only 7–9% of English’s allocation.
  • On YouTube most covered EU languages receive proportionally more moderators than English, with Spanish and Portuguese the exceptions — signaling reduced investment in Latin American user bases.
  • Global South languages average 55% of English’s per-content allocation on YouTube, falling to 7.5% on Twitter/X.
  • Interface support and moderation staffing are misaligned: some UI-supported languages (Greek, Czech, Romanian) lack dedicated moderators, while some moderated languages (Tagalog) lack UI support.

Connections

This paper is a strong empirical contribution to the DSA-driven platform transparency and data-access literature, offering a critical counterpoint to studies that assess what disclosed data can and cannot reveal — see Rieder2025-ju and Votta2025-xz on the affordances and limits of platform transparency reporting. Its focus on labour, inequity, and the governance of moderation extends classic accounts of content moderation as infrastructure such as Gillespie2010-as and Gillespie2022-jx, while its concern with who platform data and oversight serve resonates with debates on access and accountability in Gonzalez-Bailon2023-uy and Freelon2018-ao.