Algorithm Autopsy

Overview

Recommendation systems are a form of AI most students use every day without choosing to. They predict what will keep a person watching and rank content accordingly, and the app around them is designed with features like autoplay, streaks, and infinite scroll that keep people on the app longer. In this autopsy, teams examine the features of a fictional app called Loop, sort each one by what it is optimized to do, and discuss in a Socratic seminar whether a feed can shape what people believe. They finish by designing a healthier feed with specific rules, trade-offs, and a way to measure whether it works.

Objectives

• Students will explain how a recommendation system uses predictions from past behavior to rank content, and name what it is optimized for.
• Students will analyze specific engagement design features and connect each one to its likely effect on attention, emotion, or belief.
• Students will design a feed with explicit ranking rules and user controls, and justify the trade-offs using evidence from the seminar.

Step by step

1. 10–5 min

   Hook

   Why did it show me that?

   Ask students to recall the last video or post that surprised them in a feed. "Why do you think it was chosen for you?" Collect guesses on the board ("I watched something similar," "my friends liked it," "it was popular"). Say: "Every one of these guesses describes a prediction. Today we do an autopsy on a feed to see how the predictions work and what they are trying to achieve."

2. 25–13 min

   Model

   How a recommender works

   Draw a simple loop on the board: You act (watch, pause, like, share, skip) → the system records it → it predicts what you're likely to engage with next, based on your history and people with similar histories → it ranks content by those predictions → you act again. Explain that this is a kind of AI: it learns patterns from huge amounts of behavior data. The key question is what it is optimized for. A system rewarded for watch time will learn to favor whatever keeps people watching, which isn't necessarily what's true, useful, or good for them.

   Facilitator noteAvoid claiming specific platforms work in specific ways. Say "many feeds" and "can," and keep the focus on the fictional app Loop.

3. 313–27 min

   Explore

   The feature autopsy

   Teams sort the Handout A feature strips into three categories: Keeps you on the app, Shapes what you see and believe, and Gives you control. For each strip, teams must say what the feature is optimized for and who benefits. Then each team picks the one feature they think has the biggest effect on what users believe, and writes it on a sticky note for the seminar.

   Facilitator noteSome features fit more than one category. Accept a reasoned choice, and ask the team to note the second category too. Autoplay, for example, keeps you on the app and decides what you see next.

4. 427–47 min

   Debrief

   Socratic seminar: Can a feed change your mind?

   Students read Handout B, a fictional internal memo from Loop's product team, and annotate it. Run a fishbowl: the inner circle discusses for 8 minutes while the outer circle tracks evidence and listens for strong moves, then swap. Opening question: "The memo says Loop only shows people what they want. Is that true?" Follow-ups: "If a feed shows you more of what you react to, and outrage gets reactions, what happens to what you think is normal?" "Who is responsible for what you see: you, the company, the creators, or the algorithm?" "Would you trust a feed more if you could see why it chose each post?"

   Facilitator noteListen for students who distinguish showing people what they want from showing people what they react to. That distinction is the core insight. Avoid presenting "filter bubbles" as proven fact for everyone; research on how much feeds shape beliefs is ongoing, and the seminar should explore the question, not settle it.

5. 547–67 min

   Create

   Design a healthier feed

   Teams complete Handout C: they choose who the feed is for, what it is optimized for instead of pure engagement, three ranking rules, two user controls, one feature they would remove, and how they would measure whether it works. They must also name one trade-off: what the company or users would lose. Teams sketch the feed screen in the box.

   Facilitator notePush on measurement. "Users feel better" is a goal; "users report whether they learned something new after a session" is a measure.

6. 667–75 min

   Reflect

   Pitch and pledge

   Each team gives a 60-second pitch. The class votes with sticky notes on the design they'd actually use. Each student writes one sentence: "One thing I'll do differently with my own feed this week is ___."

   Facilitator noteExamples: using a "not interested" button, turning off autoplay, following a source that disagrees with them, or switching to a chronological view where available.

Adaptations

Grade 9

Use eight feature strips (1, 2, 3, 5, 7, 9, 11, 13) and run the seminar as a whole-class discussion instead of a fishbowl.

Computer science classes

Have teams write their three ranking rules as pseudocode, including how the feed would score and sort a list of sample posts.

Emergent bilinguals

Provide seminar sentence stems ("I agree with ___ because ___," "The memo says ___, but ___") and pre-teach optimize, engagement, and recommendation.

Psychology or health classes

Connect the "Keeps you on the app" features to course content on habits and rewards, and add a wellbeing measure to the design spec.

Pairs well with

• News Literacy ProjectResources on algorithms, news feeds, and how information spreads online.
• Learning Activities HubMore media and AI literacy activities, including breakouts on viral claims.