What a recommendation system does
Video sites, shopping stores, streaming services, news apps and social networks hold far more content than anyone could browse. A recommendation system selects and orders items for each person, so the feed or home page looks different for different users. The goal is generally to show things a person is likely to find relevant or interesting, which in practice usually means things they are likely to click, watch, buy or react to.
How the systems typically work
Details are proprietary and vary a great deal, but several general ideas are widely used.
- Collaborative filtering. The system finds people whose past behavior resembles yours and suggests things they liked. The logic is “people like you also enjoyed this.”
- Content-based matching. It looks at the attributes of items you engaged with, such as topic, genre or keywords, and suggests similar ones.
- Ranking models. Machine learning models predict how likely you are to take an action on each candidate item and sort accordingly. They are trained on large records of past behavior, so training data shapes what they learn.
- Context signals. Time of day, device, location and what is currently popular can influence the results.
Signals fed into these systems may include clicks, watch time, likes, shares, searches, follows and, on some services, items a person skipped. Such tracking relies on data collection, which raises its own questions about online tracking and personal data.
What is a filter bubble?
The term “filter bubble” is widely used to describe this effect. The concern is that personalization can gradually surround each person with content that matches what they already like or believe, leaving them with less exposure to different views or unfamiliar topics, and perhaps without noticing it. A related term, “echo chamber,” usually refers to social settings where people mostly hear their own views repeated, whether or not an algorithm is involved. The two ideas overlap but are not identical.
Why the concern is plausible
Several features of the system give the worry some logic:
- Ranking for engagement can favor content similar to what already held a person’s attention.
- Feedback loops are possible: what you see affects what you click, and what you click affects what you see.
- Users rarely see how or why an item was chosen.
- Pleasant, familiar material is often easier to engage with than challenging material.
Why the evidence is mixed
The research does not give a simple verdict, and it helps to hold the claim loosely.
- Some studies have found that people’s online news diets are more varied than the strong version of the bubble story suggests, and that people’s own choices, such as whom they follow and what they click, can matter at least as much as algorithms.
- Other work finds that ranking and design choices can affect what is seen, and that effects differ by platform, topic and person.
- Measuring the question is hard. Outside researchers often lack full access to platform data, results from one service may not apply to another, and services change their systems frequently.
- Offline life has its own bubbles. People often choose friends, neighborhoods and newspapers that match their outlook, so it is difficult to separate the effect of an algorithm from existing habits.
Because of this, claims that algorithms certainly divide society, and claims that they have no effect, both go further than the evidence usually allows. The wider debate about technology’s role is covered in discussions of social media and society.
Other effects beyond ideology
Narrowing is not only about politics. Recommendations can shape music taste, shopping, hobbies and the topics a person learns about. They can also reward popular items, so already-popular content tends to become more popular. Creators may adapt their work to what the system seems to favor, which can change what gets produced. Systems can encode algorithmic bias as well, for example by under-recommending some creators or groups.
What services and researchers explore
Approaches under discussion include showing a wider range of content in feeds, letting users adjust or switch off personalization, offering a simple chronological view, explaining why an item was recommended and allowing independent audits. Each has costs. Less personalization may make a service feel less useful, explanations may be hard for ordinary users to read, and audits raise privacy questions. Interest in explainable AI is partly driven by these concerns.
What individuals can do
You cannot see inside a system, but some habits are reasonable:
- Look for information from several sources, not only your feed.
- Search for topics directly, and use bookmarks or subscriptions you choose yourself.
- Check privacy settings and personalization controls where they exist.
- Occasionally follow something outside your usual interests and see what changes.
- Remember that a feed shows what the system predicts you will engage with, not an even sample of the world.
A feed is a selection. Knowing that it is selected is the first step in using it with some care.