Why it’s here
The first document filed under deployment-and-society, and the paper that
made ADR-052 add the topic — it was the highest-revisit entry on the
incoming list and the sixteen topics had no word for it.
What justifies it in a record about ML practice is not the finding about Twitter. It is the method: how to measure what a recommender does when you have no access to it, no cooperation from its operator, and no way to inspect the model. Every design decision in §3 is a reasoned trade a practitioner would face again.
The design, and the reasoning behind it
120 sock-puppets in four arms — 30 following nobody, 30 left-leaning, 30 right-leaning, 30 balanced — with the partisan arms following ten media outlets drawn from the AllSides bias chart (seven moderate, three strong) plus four political figures or party accounts.
Follow sets are controlled to the subject, not to realism. Prior audits replicate real users’ follows for ecological validity. This deliberately does not, and says why: real users follow diverse non-political accounts, which would confound a measurement about political content. The trade is stated rather than assumed.
The puppets never interact. No clicks, no follows-from-recommendations, no engagement of any kind — against the practice in radicalization studies, and for three reasons the paper gives: interaction creates feedback loops that make the baseline impossible to isolate; partisan accounts would engage differently, so cross-arm comparison stops being clean; and the target is the algorithm’s inherent bias rather than user-algorithm dynamics.
Everything the platform forces you to specify is randomized — the three required interests, birthdates between 1990 and 1999 — and a VPN removes location effects.
Exposure is weighted by rank, not counted. A “weighted occurrence per 1,000 tweets” applies an exponential decay over timeline position, calibrated from published attention data showing the top 20% of items take about 70% of views. A raw appearance count is not exposure, because nobody reads to the bottom.
Amplification is measured against the balanced arm, not left against right, so the baseline is a timeline the algorithm itself produced rather than an opposing condition.
Four collections a day, 2,000–3,000 tweets per account per day, October 2 to November 19 2024.
What was found
Exposure is concentrated everywhere. Gini above 0.45 in every arm, with all pairwise differences significant at p < 0.001. Prior work put the Gini for exposure to friends’ tweets at 0.6–0.7; out-of-network recommendation is comparably unequal.
The accounts following nobody get the most diverse recommendations. Right- leaning accounts have the highest inequality, then balanced, then left; the neutral arm is the most diverse of all, which the paper attributes to cold start. Personalization concentrates.
A handful of moderate follows is enough. Top aligned voices are amplified more than 50% above the balanced baseline in both partisan arms — from accounts following only ten media outlets and four political figures, most of them moderate. Left-arm amplification is slightly larger in magnitude than right-arm; de-amplification does not differ significantly.
The default timeline skews right. In the neutral arm, right-leaning users take 30.16% of exposure among the top 20 recommended accounts against 12.92% for left-leaning ones; 35.26% against 22.34% in the top 50; 31.39% against 20.83% in the top 100.
Conditions
One platform, one election, six weeks. Twitter/X, the 2024 US presidential election, October to November 2024. A recommender is a moving target and this is a photograph of one.
Political leaning is assigned by hand from public information, using the AllSides chart for outlets and profile descriptions plus external sources for individuals — and the paper says the classification “may be subject to inaccuracies or changes over time”. The headline asymmetries rest on it.
The decay parameters are borrowed from other platforms. The 20%/70% attention figure comes from TikTok and YouTube studies. It is a calibration assumption transplanted across platforms, and every exposure number inherits it.
Ecological validity is traded away on purpose. Non-interacting accounts with narrow, curated follow lists are not users. The paper argues the trade is right for its question and names the literature that disagrees; a reader should not take the numbers as what a person would experience.
Cold start is the explanation offered for the neutral arm’s diversity, not one that is tested. It is plausible and it is an interpretation.
And the platform’s own defaults are a confound the paper flags itself — trending topics and default settings can still skew the neutral arm that is supposed to be the unbiased baseline.