Two related ideas that are often blended
The terms “filter bubble” and “echo chamber” are frequently used as if they were one thing, but they describe different mechanisms.
- A filter bubble is the idea that personalised systems, by predicting what you will like, narrow the information you are shown so that you see less that is unfamiliar or opposing. The cause is the software.
- An echo chamber is the idea that people end up in groups, online or offline, where their views are repeated and reinforced and other views are discounted or absent. The cause is largely social: who you associate with and what you choose to trust.
The two can interact, but they are not the same, and the evidence for each may differ. Treating them as identical makes it hard to say what a study has actually shown.
Why the ideas are plausible
The appeal is easy to see. Recommendation systems optimise for predicted engagement, and people often engage more with what they already agree with or find familiar. People also tend to follow, friend and trust others who resemble them, a long-observed pattern in social life that long predates the internet. Put together, it seems natural to expect narrowing. The mechanisms are discussed in the explanation of recommendation algorithms, and the concern is part of wider debates about political polarisation.
Plausibility, however, is not evidence. The question for researchers is how much narrowing actually occurs, compared with what, and whether the platform is the cause.
How researchers study it
Studies tend to use a handful of approaches, each with limits.
Analysing public posts and networks. Researchers can map who follows or shares whom and measure how clustered the groups are. This shows connection patterns but not what people actually viewed, and people can follow accounts they disagree with.
Browsing and consumption data. Tracking records, panels or donated data can show which sources people visit. This better reflects exposure, but panels are limited samples, and they usually cover only some devices and services, an issue related to survey panels.
Surveys. Asking people about their news sources and views is feasible at scale, but it relies on self-reported behaviour, which is imperfect.
Experiments. Researchers may change what some users see, or ask participants to follow different accounts, and compare outcomes. These can address cause more directly, though they often run for short periods and may not match everyday use. Platform-run experiments are rarely available in full to outsiders. A natural change in a platform’s design can sometimes serve as a quasi-experiment, in the way described for natural experiments.
Audits with test accounts. Researchers set up accounts with scripted behaviour and record what is recommended. This isolates the system’s response but represents synthetic users, not the variety of real ones.
Why findings are mixed
Across these approaches, studies are generally described as producing a mixed picture. Several reasons explain this.
Measuring “diversity” is not straightforward. Is a mix of sources diverse if they all share the same outlook? Does exposure to an opposing view count if it is seen only briefly or in a hostile framing? Different definitions lead to different answers.
Online is only part of the diet. People get information from television, workplaces, family and friends. Studies that look at one platform may overstate or understate overall narrowness. Work on news consumption suggests that habits are spread across many sources and that many people use news only occasionally.
Individual choice and algorithmic selection are tangled. If someone sees mostly one viewpoint, is that because the system filtered it or because they clicked on it and followed those accounts? Separating the two is the core difficulty, and it is a problem of causal claims. This also involves selection effects, because the people who use a given service are not a random sample.
Groups differ. Some studies find stronger clustering among a small, highly engaged minority than among typical users. An average can hide this. Care is needed not to apply a pattern found in heavy political users to everyone, a problem close to the ecological fallacy.
Platforms and time periods differ. A design in one service and year may not describe another. Results can go out of date.
Exposure is not belief. Even if someone sees an opposing view, it may not change their mind, and some experiments have suggested that exposure to opposing content can in some circumstances be met with resistance. Seeing less diverse content is also not the same as holding more extreme views.
What can reasonably be said
Taking care not to overstate, a modest summary might be:
- There is evidence that people’s networks and sources are often clustered by shared views.
- Evidence that algorithms alone create strongly isolated bubbles for typical users is generally weaker, and many studies find that personal choice and existing habits play a large part.
- Some people, especially those deeply engaged with particular topics, may encounter narrower material, and recommendation systems may be able to reinforce that in some settings.
- Results depend on definitions, platforms and time, so confident universal statements are rarely justified.
Spaces built around a shared interest, such as online communities or fan communities, may also be internally similar by design, which is not the same as being cut off from all other information.
Reading a claim about bubbles
Useful questions include:
- Is the claim about the software, the social group or both?
- Does the evidence measure what was shown, what was clicked or what people believe?
- Who was studied, and over what period?
- Does the study compare with another source of information, or with an offline baseline?
- Could users have chosen this pattern themselves?
The idea of an information bubble is a useful prompt to ask what a feed leaves out. As an established fact about everyone’s experience, though, the evidence is more limited and varied than the term suggests.