AI projects · Build · 1-2 weeks

Teach a model with household objects, then try to break 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

You train a tiny image classifier on everyday objects (fork vs spoon, sock vs glove), then deliberately try to break it with lighting changes, odd angles, and look-alikes. The point is not a perfect model — it is to feel how brittle pattern-matching can be, and to write a clear model card for what you built.

Age band

Ages 11–14 (middle school). An adult helps with camera permission in the browser.

Time needed

1–2 weeks (several short sessions).

AI goal

Build — collect examples, train a small model, test it, and document failure modes.

Materials

Free tools only; no paid accounts required.

  • Google Teachable Machine (Image project) in a modern browser — free, no install.
  • A device camera, or 20–40 photos per class taken ahead of time.
  • Two or three object classes from around the house (examples: apple vs orange; remote vs phone; left shoe vs right shoe).
  • Notebook or sheet for a mini model card: classes, number of examples, test accuracy on held-out photos, known failures.

Optional (still free): export the Teachable Machine model and open the downloadable preview page Teachable Machine provides — no cloud account required for the basic flow.

Steps

  1. Choose classes. Pick 2–3 visually distinct objects you can photograph safely. Write a one-sentence job for the model ("Tell apart the kitchen fork and the kitchen spoon on a plain table").
  2. Collect training photos. For each class, take at least 20 photos. Vary distance a little, but keep the background boring on purpose for the first version.
  3. Hold out a test set. Before training, set aside 5 photos per class that the model will not see during training. Put them in a separate folder or album.
  4. Train. In Teachable Machine, add classes, upload or capture samples, train with the default settings, and note the on-screen training feedback.
  5. Test on held-out photos. Show each held-out photo (or point the camera at the real object in a similar setup). Record correct / incorrect. Accuracy = correct / total held-out.
  6. Break it on purpose. Try at least five attacks: dim light, bright window behind the object, object partially covered, wrong background, a look-alike object not in training. Log each attack and whether the model still looked "confident."
  7. Write a model card (half page). Include: intended use, classes, training count per class, held-out accuracy, top three failure modes, and "do not use for…" (for example, not for safety decisions).

Ethics and safety note

Do not train on photos of people's faces, classmates, or strangers. Objects only. Do not point the camera at anything unsafe. If the browser asks for camera access, an adult should approve and you should revoke access when finished.

What done looks like

CriterionNot yetDone
DataOne class, or under 15 images each2–3 classes, ≥20 train images each, 5 held-out each
EvaluationOnly live webcam vibesHeld-out accuracy written as a fraction
Failure log"It worked pretty well"≥5 deliberate break attempts with outcomes
Model cardMissing limitsIntended use + failures + do-not-use listed

Adaptations

Classroom

Stations share devices. Each group trains a different object pair and swaps break-it tests.

Homeschool

Use toys or pantry items. Compare a "clean background" model to a "messy table" remake.

Weekend

Use 2 classes only and 15+15 training images if time is tight; keep the held-out set.

Research

Change one variable (background clutter) and retrain; compare held-out accuracy before vs after.

Science fair

Demo live break-it attempts next to the model card; judges should see a failure, not only a success.

All AI projects

Teach a model with household objects, then tr… · Schoolhouse