Every time you open a social app, a ranking system chooses a few dozen posts from thousands of candidates and puts them in a specific order.
Political content goes through the same selection as everything else. Our earlier piece on how social media shapes voter opinions covered bots and filtered politics.
This one covers the ranking system underneath, who controls it, why nobody outside the company can check it, and how to audit your social media feed for political bias with a method you can run yourself.
The ranking decision no voter ever sees
A feed is the output of a sorting process. The platform gathers candidate posts from accounts you follow, accounts you do not follow, groups, and ads.
A model then scores each candidate on how likely you are to react to it, and the highest scores appear first.
People at the company decide which reactions count and how much each is worth. A reply may be weighted more than a like, a share more than a reply, and time spent looking at a post may count too.
Those weights change, usually without notice to users, and the change is visible to you only as a feed that feels different.
A voter sees the result and never the decision. Two people who follow the same accounts can see different orders, and neither can find out why.
For a system that may sit between a candidate and a voter, that is an unusual level of opacity.
What a recommendation engine optimizes for, and why it is not civic information
Recommendation engines are built to predict engagement, because engagement is what the business sells, time on the platform and attention for advertisers.
Accuracy, relevance to a ballot decision, and usefulness to a citizen are not what the model is trained to predict.
One documented case shows the consequences. In 2018 Facebook changed its feed to emphasize “meaningful social interactions.”
Internal documents disclosed in 2021 and reported by the Wall Street Journal showed employees warning that the change rewarded divisive and outrage-driven content.
The company’s stated goal was healthier interaction. The measured behavior of the model was something else.
Civic information performs poorly against that objective. A post explaining where and when to vote competes for placement against posts engineered to provoke a reaction.
A dry, accurate explainer on a ballot measure is unlikely to outscore a furious clip, because the model rewards the reaction and has no concept of what the post is for.
Why outsiders cannot check the claims
Three barriers keep independent verification out of reach.
Closed data. The ranking model, its training data, and what it showed any individual user at any moment are held inside the company. Outsiders see only what the platform chooses to publish.
Limited researcher access. Tools that let outside researchers study public content have been cut back. Meta shut down CrowdTangle in August 2024 and replaced it with the Meta Content Library, which has an application process and narrower access. X ended free access to its API in 2023. Researchers who want to study political amplification now face higher costs and more gatekeeping.
Self-reported fixes. When a platform says it has reduced political content or fixed an amplification problem, the announcement is usually the only evidence. Transparency reports are written by the company about itself. In 2023 Twitter released portions of its recommendation code, but code without the underlying data and the live state of the system does not let anyone reproduce what a particular person was shown.
The consequence is that every claim about platform fairness, from critics and from companies alike, rests on partial evidence.
A company can say its system is neutral and nobody can test it. A critic can say it is biased and face the same limit.
Where platform influence on voters is documented and where it is still only suspected
Some effects are documented in peer-reviewed work:
- Turnout messages. A study published in Nature in 2012 followed a 2010 Facebook experiment on about 61 million users. People shown a message with the faces of friends who had voted were more likely to vote, and the authors estimated the message produced roughly 340,000 additional votes.
- Emotional tone. A 2014 study in PNAS altered the mix of positive and negative posts in the feeds of 689,003 Facebook users. The effect on what people then posted was small but measurable, which showed that feed composition can shift behavior.
- Algorithmic amplification. A 2022 PNAS study by researchers working with Twitter found that in six of seven countries studied, the algorithmic timeline amplified posts from mainstream right-leaning politicians more than the reverse-chronological timeline did. In the U.S. it also amplified right-leaning news sources relative to left-leaning ones. This is a measurement of amplification, not of changed votes.
Other claims remain suspected. In 2023, researchers working with Meta published studies on the 2020 election in Science and Nature.
One experiment switched users to a chronological feed. It changed what they saw and how long they stayed, but over the study period it did not produce a detectable change in polarization or political attitudes.
That is a real finding, and it complicates the simple claim that the algorithm alone polarizes voters.
It does not settle the question. The studies ran for a limited time, on users who consented, with Meta’s cooperation and Meta’s data.
Critics have also pointed out that the platform’s own settings during that period may have differed from normal conditions. Whether ranking systems shift vote choice across a whole electorate has not been shown, and has not been ruled out.
What real transparency would look like
Meaningful oversight would have specific parts, and each can be written into a rule.
- Independent audits with real access. Auditors outside the company would be able to run tests against the live system and examine data, under privacy protections, not just read a summary.
- Data access for vetted researchers. Qualified academics and journalists would have a legal right to platform data on political content, not access that depends on the company’s goodwill.
- Plain-language disclosure of political ranking. Each platform would state in ordinary words whether it boosts, reduces, or ignores political content, how it defines “political,” and when that policy changed. Meta announced in 2024 that Instagram and Threads would stop recommending political content from accounts users do not follow, unless users opt in. Without disclosure and outside verification, a user cannot confirm what that policy does in practice.
- Change logs near elections. Any alteration to ranking weights in the weeks before an election would be recorded and made available.
- A non-personalized option. Users would be able to choose a feed not tailored by profiling.
The European Union’s Digital Services Act already moves in this direction for very large platforms.
It includes provisions for vetted researcher access and requires a recommendation option that does not rely on profiling. The U.S. has seen proposals for similar access, but at the time of writing no comparable federal law is in force, which leaves access up to the companies.
How to audit your social media feed for political bias before the next election
You cannot inspect the model, but you can measure what it shows you.
A personal audit documents your own feed. It cannot show why the feed looks as it does, and it does not prove anything about other users. It gives you a record to look at instead of an impression.
1. Set the rules before you count. Decide your categories in advance, political or not political, source type (news outlet, politician or campaign, ordinary user, ad), and whether you follow the account or the post was recommended. Judge posts by these categories and not by whether you agree with them.
2. Record the first 50 posts. Open the app without searching or interacting. Log each post as you scroll. Do this on three different days and at different times.
3. Compare feed modes. Most platforms offer a “Following” or chronological view next to the default one. Log 50 posts from each. The difference between them shows how much the ranking system adds to what your own follows would give you.
4. Count the recommended share. Tally how many posts come from accounts you do not follow. Note what proportion of those is political and what sort of tone they have, such as outrage, alarm, or ridicule.
5. Check the ads. Meta’s Ad Library and Google’s Ads Transparency Center list political and issue ads. Search for the candidates and measures on your ballot and compare the ads you have seen with what is archived.
6. Look at what the platform says about you. Review your ad preferences and any “why am I seeing this” explanations. Download your account data from the platform’s settings and see which interests and advertiser lists it has assigned to you.
7. Test with a second view. Ask a friend with different habits to run the same count, or open the platform logged out. Differences tell you how much of your feed is shaped by your own history.
When you finish, compare the numbers. Look at the share of political posts, the balance of sources, the share of recommended content, and the emotional tone. A feed dominated by recommended posts of one tone is worth noting, and the notes are the evidence to take to friends, local reporters, or a representative’s office.
Run your audit this week
Spend 20 minutes this week running the feed audit above, then confirm your registration using our safe voter registration guide. Registration deadlines and rules differ by state, so check yours before the November midterms.
Then leave a comment below and tell us what your feed audit turned up.