How Algorithms Shape What You See

What a feed actually is

When you open a social platform, you are not seeing everything posted by the accounts you follow, in the order they posted it. You are seeing a ranked list. A system scored a large pool of candidate items and selected the ones it predicted you were most likely to respond to, in an order it predicted would keep you responding.

This is worth stating plainly because the visual design of a feed suggests something different. An infinite vertical list reads like a record of what is happening. It is closer to a personalized recommendation, in the way that a shop’s front table is a recommendation rather than an inventory.

The distinction matters because most people use the feed as evidence about the world beyond it. This is not an unreasonable thing to do, but it requires an adjustment, and the adjustment is what this lesson provides.

What these systems optimize for

Ranking systems differ across platforms and change frequently, so specific claims about any one system date quickly. What is stable is the general shape.

A ranking system predicts the probability that you will take some action on an item: read it, watch most of it, react, reply, or forward it. It combines those predicted probabilities into a score, weighted by what the platform values, and sorts by the result.

Three consequences follow directly from that structure, and they hold regardless of the specific weights.

Predicted reaction stands in for value. The system cannot measure whether something is true, important, or good for you. It can measure whether people like you engaged with things like it. Everything else has to be approximated by that.

Content that provokes response is advantaged. Not because anyone chose to promote outrage, but because response is the measurable quantity. Material that produces a strong reaction produces more of the signal the system is scoring.

Your history narrows your future inputs. Each item you engage with is training data for what you are shown next. This does not require a conspiracy or a filter bubble in the strong sense. It is the ordinary behaviour of a feedback loop.

Three inferences to stop making

“This is popular, so many people believe it”

What you can conclude from something appearing repeatedly in your feed is that the system predicted you would engage with it. Repetition within your feed is weak evidence about prevalence in the population, because the selection was made for you specifically.

Vosoughi and colleagues, examining roughly 126,000 story cascades shared by about 3 million people on Twitter between 2006 and 2017, found that false stories diffused significantly further, faster, and more broadly than true ones, and that this held after accounting for account characteristics such as follower count and verification status [2]. Reach, in other words, is not a proxy for accuracy.

“Everyone is talking about this”

“Everyone” here means the subset of accounts the system selected, filtered further by who chose to post publicly. Silence is invisible in a feed. So is disagreement expressed by scrolling past.

“This is urgent because it is at the top”

Position reflects predicted engagement, not importance or recency. A post from three days ago can appear above a post from ten minutes ago. Where a genuine emergency is underway, official channels and local emergency services are the appropriate source, not feed position.

What to do instead

Read the source, not the surface. Open the original item rather than judging from the summary card. This costs a few seconds and removes a whole class of errors, including quote cards that misattribute, headlines that overstate their own article, and screenshots that crop.

Check prevalence outside the feed. If a claim is about how common something is, the feed is close to the worst available instrument. Look for a survey, an administrative dataset, or a report that states its sample and method.

Change one input deliberately. Follow two or three sources you would not otherwise see, and note over a week whether your sense of what is happening shifts. If it shifts substantially, that tells you something about how much of your prior picture was selection.

Use chronological or list views when available. Many platforms retain an unranked view. Comparing it with the ranked view for a few minutes is the fastest way to see the selection at work.

Context worth having

Reliance on these systems for news is now the majority pattern rather than a minority one. The Reuters Institute’s 2026 survey of 48 markets found that 54 percent of respondents used social media and video networks as a source of news, rising to 56 percent when AI chatbots are included, and that 10 percent reported using AI chatbots for news, up from 7 percent the previous year [1].

That does not by itself mean people are badly informed. It means the intermediary between events and audiences is increasingly a ranking system rather than an editor, and that reading skill has to account for it.

Practical exercise

Spend ten minutes with one feed and record three things for each of the first ten items:

1. What is the item, in five words?
2. What signal do you think caused it to be selected for you?
3. If you had to estimate how many people outside your feed hold the view it expresses, what would you say, and what would you base that on?

The third column is the important one. Most people find they cannot answer it, which is the point.

Summary checklist

• I know a feed is a ranked selection, not a sample.
• I do not read frequency in my feed as prevalence in the world.
• I open originals rather than judging from summary cards.
• I check claims about how common something is against a source that states its method.
• I have compared a ranked and an unranked view of the same feed at least once.

References

Numbered citations in the text above correspond to the entries below.

1. Reuters Institute for the Study of Journalism. (2026). Digital News Report 2026: Overview and key findings. University of Oxford. https://reutersinstitute.politics.ox.ac.uk/digital-news-report/2026/dnr-executive-summary
2. Vosoughi, S., Roy, D., & Aral, S. (2018). The spread of true and false news online. Science, 359(6380), 1146-1151. https://doi.org/10.1126/science.aap9559

Next step

Audit your own feed with the worksheet

Topics covered: Digital LiteracyInformation Manipulation