Internet Censorship Course
Filter bubbles are created when algorithms personalize content for users based on their past behavior and preferences. This can lead to users only seeing content that is tailored to their interests, which can limit their exposure to new ideas and perspectives. This can lead to an echo chamber effect, where users are only exposed to content that reinforces their existing beliefs and opinions.
In this activity, you will try to uncover evidence of filter bubbles on online search portals (search engines, news portals, etc.).
Search engines often use algorithms to personalize search results for users. Personalization algorithms are based on a variety of factors, including user geography, search history, and past behavior, the collection of which constitutes a user “profile”. This profile is used to tailor search results to a user’s preferences. It may also be used to target advertising to a user.
A potential drawback of personalization is that it potentially results in users only seeing content that match their past (or existing) interests.
Try the following experiment. You may want to work with a partner for this exercise, as different users have different profiles and may thus see different search results.
Pick a handful of search terms that you are interested in. Choose a range of topics, from politics, to current events, to retail, to local businesses.
What do you notice about the differences in the search results? Are there any differences that you find surprising? Are any aspects of Google’s that you could imagine are tailored based on your past behavior on Google products? (Note that DuckDuckGo does personalize results based on your current profile, such as your geography, but does not perform tracking or history.)
A search engine at least waits for a query. A feed (TikTok’s For You page, Instagram Reels, YouTube Shorts, the X “For You” timeline) decides what you see with no question asked, from signals you may not know you are giving: what you watched to the end, what you scrolled past, how long you paused. There is no query to vary, so the thing to probe is the account itself. This part asks you to audit the “algorithm” the way an outside researcher would: from its outputs, with controlled inputs.
Work with a partner so you have two feeds to compare. Use your own accounts, or a logged-out or fresh session where the platform allows it; do not engage with content you would not want in your history.
Baseline. Each of you opens the same feed and scrolls the next 20 to 30 items without liking, following, or watching anything to the end. Tally what you were shown by rough category (news, politics, sports, music, shopping, a hobby, and so on). How different are the two feeds already, and what in your histories would explain it?
Intervene. Pick one topic that is absent from your baseline and harmless to you (a sport you do not follow, a cooking style, a city). For five minutes, engage only with that topic: watch those items to the end, like a few, skip everything else quickly. Then scroll another 20 to 30 items and tally again. How many items did it take for the feed to move? How much of the feed is now that topic?
Try to get out. Use whatever the platform offers (not interested, hide, clear watch history, “reset” or “refresh” the feed, switching to a chronological or Following view). Scroll again. Did the topic fade, and how quickly? This is the question the class kept coming back to: how much agency does a user actually have?
Compare notes. Put the two feeds side by side after the same intervention. Did the same five minutes move you both the same way? Then find what the platform says the feed optimizes for (its “how recommendations work” page) and compare it with what you observed. Does personalizing for engagement look different from personalizing to help you find things, and could you tell the two apart from the outside?
Keep your tallies. They are small, but they are the same kind of evidence an audit study collects, and they show both what you can conclude from the outside and what you cannot: you saw the outputs move, but not which signal moved them.
Increasingly, people ask a chatbot instead of a search engine, and the chatbot returns one synthesized answer rather than a list of links. That makes the question from Part 1 harder to answer, not easier: there is no result page to compare, and nothing tells you what the model left out.
Repeat the experiment with two or three assistants (ChatGPT, Claude, Gemini, Copilot, Perplexity, or a model served from another country).
Reuse a few of your Part 1 queries (and, if you did Part 2, the topic you pushed your feed toward), especially the political or current-events ones and anything with a regional angle (an election, a border dispute, a local health question). Ask each model the same question, worded the same way.
Vary the vantage point the way you did for search: logged in versus out, a fresh session, and, if you have a VPN, the same question asked from two or three different countries. Try asking in a second language if you can.
Compare the answers side by side. Which sources does each model cite, if any? Does the framing change with location or language? Does one model decline or hedge where another answers directly? Do two sessions of the same model agree with each other?
Note what you can and cannot conclude. A search engine’s personalization can be measured by diffing result lists from many vantage points, which is how the studies in the lecture work. For a chatbot, what would the equivalent measurement even be? Think about what you would need to control, and how you would tell personalization from randomness in the model’s output.
Bring your tallies from Parts 1 to 3 and work through these in your group. Several of them came straight out of class discussion; none has a settled answer.
Is personalization censorship?
Looking versus being fed
Agency: the algorithm, or us?
Transparency and regulation