Source
Ye, Luceri and Ferrara (2024), LIT-480 — read as NOTE-229.
What it explains
| practice | what it says to do | what this says it is |
|---|---|---|
| SOTA-290 | include an arm that follows nobody | not a control that reports the null, but the arm that carries the finding — the ordering across arms is the measurement |
The account
The measurement is LIT-480’s sock-puppet audit of Twitter/X’s “For You” timeline: 120 accounts in four arms of 30 that never interacted, four timeline collections a day for six weeks, 9.79M tweets, with exposure weighted by rank.
The audit ran four arms: accounts following nobody, accounts following ten moderate-to-strong media outlets and four political figures on the left, the same on the right, and a balanced set. Concentration of exposure was measured by Gini coefficient over the accounts appearing in each timeline.
Every arm is concentrated — Gini above 0.45 throughout, comparable to published figures of 0.6–0.7 for exposure to one’s own friends. The ordering is the result:
right-leaning most concentrated > balanced > left-leaning > following nobody, least concentrated of all.
All pairwise differences significant at p < 0.001.
So the arm with no personalization signal at all sees the widest range of sources, and every arm that supplies a signal sees a narrower one. The direction is the opposite of the intuition that a recommender broadens what you would otherwise have found: on this measurement personalization is a narrowing operation, and the baseline it narrows from is the untargeted feed.
And the amount of signal required is small. The partisan arms follow ten media outlets — seven of them moderate — and four political accounts. That is enough for the top aligned voices to run more than 50% above the balanced baseline. No interaction, no clicks, no history: only the follows, and only a fortnight of them before the measurement window.
The source’s own reading is that once a new user follows a few partisan accounts, their recommendations fill with like-minded voices quickly. The measurement supports the speed as well as the direction, because the arms had no time to accumulate behaviour — there was none to accumulate.
Why Proposed
One platform, one six-week window, one election. The ordering is clean and significant and it is a photograph. A ranking system is changed continuously, and this one was changed by its owner during the period it was measured.
The mechanism for the neutral arm is offered, not tested. Cold start is the explanation given for why accounts following nobody see the most diverse timeline — plausible, and an interpretation. A different explanation, that untargeted feeds are simply drawn from a broader popularity pool, would predict the same ordering.
And the control arm is not clean, which the source says about its own design: accounts following nobody still receive trending topics and platform defaults. The arm that anchors the ordering is the one with a known confound.
The promotion condition therefore asks for a second platform rather than more of this one. If the ordering holds elsewhere, “personalization concentrates” is a property of the class; if it does not, it is a fact about Twitter/X in late 2024.
What this does not say
It does not say concentration is caused by personalization. The arms differ in what they follow, and following a set of accounts changes both the personalization signal and the candidate pool. Untangling those needs an arm that follows many accounts chosen to be uninformative, which was not run.
It does not say the concentration is harmful. A concentrated feed of things a reader wants is a working recommender by most operators’ definition. What the measurement establishes is the direction and the magnitude, and the record files that without the evaluative step.
It does not carry the partisan asymmetries. The finding that default timelines skew right, and that left-arm amplification is slightly larger in magnitude, rests on hand-assigned political leanings the authors describe as possibly inaccurate or stale. That is a separate claim with a separate evidential basis, and this account does not depend on it.