Skip to content
Y4: AI Science-Fair Project (Capstone)Lesson 9 of 12

9 · Bias & limitations check

Great scientists don't just report their score — they tell you what their project's limits are, and they check it for bias. Doing this makes your project more trustworthy, not less, because it shows you understand your own work.

Check for bias (remember F3 and Code.org). Bias is when a model leans one way unfairly because of its examples. Ask yourself:

  • Were my examples balanced? If "ripe" had 100 photos and "unripe" had only 10, the model learned one class way better than the other. That's a kind of bias — fix it by balancing your examples.
  • Were my examples varied enough? If every "dog" photo was a golden retriever, the model might fail on a poodle. It learned "golden retriever," not "dog."
  • Could my model be unfair if it touched people? This is exactly why we studied objects and our own poses. Code.org's "How AI Works" covers algorithmic bias for students — the same idea you met in F3.

Name your limitations (every honest project has them). Limitations are the things your project couldn't do or check — and saying them out loud is a strength:

  • "I only tested 20 photos — a bigger test might show a different score."
  • "All my photos were taken in my kitchen — it might not work in different lighting."
  • "I only used 2 classes — real fruit ripeness is more of a slow change than two boxes."

Did AI help? Say exactly how — that's honesty in action. If an AI tool helped you (brainstorming, explaining an idea, helping organize data), write one honest line about it, e.g. "I used an AI helper to suggest project ideas; I chose and tested the question myself and checked all facts in trusted sources." UNICEF and good fair rules want AI used transparently — and you already know from F3 that being open about how you used a tool is the brave, right choice.

A project that says "here's what worked, here's what didn't, and here's how I used AI" is more convincing than one that pretends to be perfect.

Think about it. Name one possible bias in your example data and one real limitation of your project. How could a future version fix each one?

Sources

Sign in to track your progress