fg-zettelkasten

Visual identities in troll farms: The Twitter Moderation Research Consortium

3 min read

Visual identities in troll farms: The Twitter Moderation Research Consortium

Summary

This paper presents a ten-year longitudinal audit of how documented changes to Facebook’s News Feed ranking algorithm shaped engagement with news content from The Guardian between 2011 and 2020. By combining a hand-curated timeline of 52 algorithmic updates with roughly one million Guardian articles and matched CrowdTangle engagement data, Bastos shows that ranking interventions produce measurable, lagged effects on engagement — but only for hard news, not for Opinion, Lifestyle, Arts, or Sport. The argument is twofold: empirically, recommender systems are not inscrutable “black boxes” but auditable sociotechnical assemblages; politically, the EU Digital Services Act (especially Article 40(4)) provides the legal scaffolding to institutionalise such longitudinal independent audits.

(Note: the supplied title refers to troll farms and the Twitter Moderation Research Consortium, but the structured summary describes a different study on Facebook News Feed and The Guardian. The note below follows the structured summary.)

Key Contributions

  • First decade-long empirical model of Facebook News Feed ranking effects on news engagement, segmented by content type.
  • A replicable methodological pipeline combining documented algorithm-change timelines with cross-correlation, Granger causality, and anomaly detection.
  • Empirical evidence that platforms apply differentiated algorithmic treatment to hard versus soft news.
  • Reframes recommender systems as auditable institutional artefacts rather than opaque black boxes.
  • Policy argument for operationalising DSA Article 40(4) and Recital 85 to enable sustained independent algorithmic audits.

Methods

  • Compiled a timeline of 52 News Feed ranking updates (2011–2020), coded by impact level and valence toward trusted news.
  • Harvested ~1,020,163 Guardian articles via the Guardian Open Platform API across five sections (News, Opinion, Sport, Arts, Lifestyle).
  • Retrieved CrowdTangle engagement metrics (likes, shares, comments, Reactions) for 576,673 matched URLs.
  • Log-transformed series and verified stationarity with Augmented Dickey–Fuller tests.
  • Applied cross-correlation function (CCF) analysis to identify lags, then Granger causality tests at those lags, plus Seasonal Hybrid ESD anomaly detection. Daily aggregation yielded 20,075 observations; analyses in R.

Findings

  • Optimal lags between ranking changes and engagement effects were 19–24 days, consistent with phased platform roll-outs.
  • News (ACF .35) and Sport (.31) showed the strongest cross-correlations; Opinion, Lifestyle, Arts were weaker (.27, .24, .25).
  • Granger causality was significant only for the News section (F = 1.6774, p = .0327 at lag 19), isolating hard news as the category most shaped by ranking interventions.
  • Anomaly clusters in 2014–2016, peaking mid-2016, align with the well-known pivot toward friends-and-family content.
  • Sport anomalies in July 2014 reflect the FIFA World Cup, not algorithm changes — a useful negative control.
  • Engagement rose steadily through 2016 then declined as major ranking updates rolled out.

Connections

This paper sits squarely in the platform-governance and algorithmic-auditing strand of the register, complementing work on independent external observation of recommender systems and platform data access regimes — see Rieder2026-pp, Rieder2025-ju, and Ohme2026-nv on auditing and access infrastructures, and Helmond2026-ll on platforms as institutional/sociotechnical objects. Its longitudinal CrowdTangle-based design also resonates with methodological reflections on platform data in Bruns2025-fz and Bastos2025-ya, while its concern with how algorithmic gatekeeping reshapes news distribution connects to Hurcombe2025-cs and Munger2025-cz.

Graph View

Backlinks