AI projects · Build · 1-2 weeks
Build a tiny recommender and audit it
These projects are not aligned to any state standards or curriculum. They are generally about learning, understanding, using, and evaluating AI.
Ages 11-14
Overview
Recommendation systems suggest "what next" using simple scores. You will build a tiny, transparent recommender for a toy catalog (books, games, or science-fair ideas), then audit who gets ignored by the ranking rule you chose.
Age band
Ages 11–14 (middle school). Needs comfort with averages and ranking.
Time needed
1–2 weeks.
AI goal
Build — implement a simple ranking rule, then evaluate its side effects.
Materials
Free tools only; no paid accounts required.
- A spreadsheet (Google Sheets or LibreOffice Calc) or a short Python script in a free browser notebook (for example the free tier of a public notebook host an adult approves). No paid API keys.
- A toy catalog of at least 20 items you invent or copy from a public domain list. Each item needs: id, title, tags (2–4), and a popularity count you assign for the simulation.
- Paper for an audit table.
Do not scrape private user data. Invent synthetic users (U1–U8) with preference tags.
Steps
- Define the job. Example: "Given a user's preferred tags, return the top 5 catalog items."
- Build the catalog table. 20+ rows. Keep tags consistent (spelling matters).
- Invent eight synthetic users with 1–3 preferred tags each. Diversify: some rare tags, some common.
- Pick a ranking rule (choose one and write it in math words):
- Score = (number of overlapping tags) × 10 + log-friendly popularity term you define, or
- Score = overlap count only (ignore popularity), or
- Score = popularity only (ignore overlap) — useful as a bad baseline
- Compute top-5 for each user. Show the score arithmetic for at least two users by hand so the sheet is trustworthy.
- Audit.
- Coverage: how many catalog items never appear in any top-5?
- Popularity bias: average popularity of recommended items vs catalog average
- Tag desert: which tags never trigger a recommendation?
- Fix attempt. Change one rule (for example add a small bonus for never-recommended items). Re-run for all users. Report what improved and what got worse.
- Write a one-page system card: inputs, score formula, audit metrics, known harms (example: niche items buried).
Ethics and safety note
This is a toy. Do not deploy it on real classmates' private preferences without consent. Do not use sensitive tags (health, religion, immigration status). Keep the catalog playful and school-safe.
What done looks like
| Criterion | Not yet | Done |
|---|---|---|
| Catalog | Under 20 items or missing tags | ≥20 items with tags + popularity |
| Transparency | Black-box "the AI ranked it" | Written score formula + hand-checked examples |
| Audit | Only accuracy vibes | Coverage, popularity bias, tag desert reported |
| Revision | No second run | One rule change with before/after notes |
Adaptations
Classroom
Groups use the same catalog but different score rules, then compare audits.
Homeschool
Use a bookshelf inventory as the catalog if you prefer real titles you own.
Weekend
12 items and 4 users if needed; keep the audit metrics.
Research
Prove a tiny inequality: if score ignores overlap, users with rare tags get lower average overlap in top-5 — show the averages.
Science fair
Live spreadsheet demo: change a tag and watch rankings move.