What is algorithmic amplification?
Algorithmic amplification is the extra reach a ranking algorithm gives a message beyond what a plain chronological feed would give it. It stopped being a matter of accusation in December 2021, when Twitter measured it on itself and published the result: in six of the seven countries studied, tweets from the mainstream political right were amplified more than tweets from the mainstream political left. No regulator forced the study and no whistleblower carried it out. The platform ran the experiment on its own users, and its own researchers put their names on the finding.
Primary source 01 / 08
Content on Twitter’s home timeline is selected and ordered by personalization algorithms. By consistently ranking certain content higher, these algorithms may amplify some messages while reducing the visibility of others.
The control group only a platform could build
Every outside audit of a ranking system hits the same wall: you cannot observe what the feed would have shown without the algorithm. Twitter had the counterfactual running in production. When it switched the home timeline to machine-learned ranking in 2016, it held back a random slice of accounts that kept the old reverse-chronological feed, and left them there:
Primary source 02 / 08
We provide quantitative evidence from a long-running, massive-scale randomized experiment on the Twitter platform that committed a randomized control group including nearly 2 million daily active accounts to a reverse-chronological content feed free of algorithmic personalization.
Against that baseline the researchers measured the reach of 3,634 accounts belonging to elected legislators in seven countries: the United States, Japan, the United Kingdom, France, Spain, Canada, and Germany. Amplification of 0 percent means ranking and chronology reach the same share of an audience; 100 percent means the ranked feed doubles it.
Six of seven countries, one direction
Primary source 03 / 08
Our results reveal a remarkably consistent trend: In six out of seven countries studied, the mainstream political right enjoys higher algorithmic amplification than the mainstream political left.
The direction was consistent and the gaps were not small:
Primary source 04 / 08
With the exception of Germany, we find a statistically significant difference favoring the political right wing. This effect is strongest in Canada (Liberals 43% vs. Conservatives 167%) and the United Kingdom (Labor 112% vs. Conservatives 176%).
An amplification of 167 percent means the ranked feed put those tweets in front of an audience more than two and a half times the size the chronological feed reached. The pattern held when the researchers moved from politicians to the press, using two independent media-bias ratings to classify millions of shared news links:
Primary source 05 / 08
Consistent with this overall trend, our second set of findings studying the US media landscape revealed that algorithmic amplification favors right-leaning news sources.
Primary source 06 / 08
The personalization algorithms amplify sources that are more partisan compared to ones rated as Center.
Read that second finding next to the first. The engine did not only lean; it rewarded the loudest versions of each side over the outlets rated closest to center, under both rating systems the study checked.
What the study did not find
Two popular beliefs failed the measurement. The first is that the algorithm pushes people toward the extremes:
Primary source 07 / 08
We further looked at whether algorithms amplify far-left and far-right political groups more than moderate ones; contrary to prevailing public belief, we did not find evidence to support this hypothesis.
The second is that the tilt operates account by account. At the level of individual politicians it dissolved:
Primary source 08 / 08
While tweets from some individual politicians are amplified up to 400%, for others, amplification is below 0%, meaning they reach fewer users on ranked timelines than they do on chronological ones.
A permutation test found no statistically significant link between an individual politician's party and their amplification. The tilt is a property of the parties as groups, not a bonus paid to every account on one side.
A measurement, not a motive
The paper stops where its data stops. It does not claim the ranking model was built to favor anyone, and the authors wrote that the causal mechanism behind the disparity invites further study. What it establishes is narrower and harder to wave off: engagement ranking is not politically neutral in its effects, and that is the platform's own arithmetic, run on its own users with impression data no outside auditor can reach. Five of the six authors were or had been Twitter employees while the work was carried out; the sixth was a paid consultant for the company.
This chapter is the measured counterpart to the testimony around it. Frances Haugen testified that engagement ranking amplifies division; this study is a number on the political direction, from inside the machine. The attention economy explains why the model optimizes for engagement in the first place. The next chapter turns the same data machinery around: instead of ranking what everyone might see, it profiles you and picks the message only you will.