Virtual Coral Reef Labs: Bringing Marine Science to Inland Schools

Lesson Overview

Time required: 2-3 class periods (45-50 minutes each)

Learning objectives: - Students understand what AI bias is - Students can identify how bias enters machine learning - Students can recognize bias in algorithmic outputs - Students can propose approaches to reducing bias

Materials needed: - Computer with internet access (one per small group or shared) - Printed worksheets (provided below) - Access to Teachable Machine (google.com/teachableml) or ML4Kids

Standard alignment: NGSS 6-8 practices include "analyzing and interpreting data" and "constructing explanations"

Lesson Plan

Part 1: Introduction to Bias (10 minutes)

Teacher introduction: "Have you ever noticed that a tool or system seems to work differently for different people? Maybe a video recommendation system recommends different things to different people. Maybe a photo app recognizes faces better for some people than others. That's often bias. Algorithms can be biased even when the person who created them wasn't trying to be biased."

Discussion prompt: "Can you think of examples where a technology seemed unfair to some people?"

Let students share examples. Common ones: facial recognition not recognizing darker skin tones, voice recognition working better for some accents, recommendation algorithms recommending different content based on demographics.

"Today, we're going to see how bias happens in AI. We're going to create our own algorithm and see if it's biased."

Part 2: Understanding Training Data (10 minutes)

Present a scenario: "Imagine we're training an algorithm to identify different types of plants. We show it pictures of 100 plants to learn from. But 80 of those pictures are sunflowers, and only 10 are roses, and 10 are dandelions."

Discussion prompt: "If the algorithm saw mostly sunflowers in training, what will happen when we show it a new rose?"

Guide students to see that the algorithm learned mostly about sunflowers. It might not be very good at recognizing roses because it had few examples to learn from.

"This is one way bias enters AI. If the training data is skewed—if some examples are much more common than others—the algorithm learns better from the common examples and worse from the uncommon ones."

Part 3: Hands-On Activity—Train an Algorithm (30 minutes)

Setup: Use Teachable Machine or ML4Kids to have students train a simple classification algorithm.

Activity flow:

1. Choose a classification task. Good options for middle school: - Classify facial expressions (happy/sad/surprised) - Classify hand gestures (thumbs up/down/peace sign) - Classify drawing types (circle/triangle/square)

2. Collect training data with balanced groups. Have students divide into small groups. Each group trains on the same task. - "When we train on happy faces, we need examples from different people. Different ages, different ethnicities, different genders. Let's make sure our training data includes everyone." - Collect examples. Take photos or drawings from students.

3. Deliberately create imbalanced training data. This is key. - Have Group A train with balanced data (equal examples of each type, from diverse people) - Have Group B train with imbalanced data (mostly one type, mostly from one demographic) - "Let's see what happens with biased training data."

4. Test the algorithms. Use new images/drawings that the algorithm hasn't seen. - Group A's algorithm (trained on balanced data) should be accurate across groups - Group B's algorithm (trained on biased data) should be less accurate for underrepresented groups

5. Compare results. Show both groups' results. - "Which algorithm is more accurate overall?" - "Is one algorithm more accurate for some people than others?" - "Why did this happen?"

Part 4: Discover Bias (10-15 minutes)

Guided discovery:

1. Look at accuracy by demographic group. Disaggregate results. - "When we test Group B's algorithm on people from the majority group, how accurate is it?" - "When we test it on people from underrepresented groups, how accurate is it?" - Make the disparity visible.

2. Trace bias to training data. Help students connect the observation to the cause. - "Group B's algorithm was less accurate for some people. Why?" - "Which group was more represented in the training data?" - "So the algorithm learned better from that group."

3. Name the bias. Use appropriate terminology. - "This is an example of representation bias. When some groups are underrepresented in training data, the algorithm learns worse patterns for those groups."

Part 5: Discussion—Why Does Bias Matter? (10 minutes)

Discussion prompts: - "If we use this biased algorithm in the real world, what could go wrong?" - "Who would be disadvantaged by an algorithm that's more accurate for some people than others?" - "Why might someone not notice this bias?"

Lead students to understand that when algorithms are used to make decisions (hiring, lending, criminal justice, healthcare), bias leads to unfair treatment.

Part 6: Proposing Solutions (10 minutes)

Ask: "If you were redesigning this algorithm to reduce bias, what would you do?"

Let students brainstorm. Common answers: - "Include more examples from all groups in training data" - "Test the algorithm on all groups to find bias" - "Use a different algorithm that's less prone to this bias" - "Have people review the algorithm's decisions"

Affirm good ideas and discuss tradeoffs. For example, including more training data takes time and effort. Testing on more groups requires more data. This is realistic—reducing bias requires work.

Worksheets and Assessments

Worksheet 1: Observation (during Part 3-4) Our algorithm was trained on: [Students describe training data]

When we tested it: - Accuracy on [Group 1]: [percentage] - Accuracy on [Group 2]: [percentage] - Accuracy on [Group 3]: [percentage]

We noticed: [Students describe what they observed about accuracy]

Worksheet 2: Analysis (during Part 5) Why might this bias matter? [Students explain consequences]

Who could be harmed? [Students think about real-world impacts]

What would you do differently? [Students propose solutions]

Assessment: - Can students identify that bias existed? (Look for: "The algorithm was more accurate for some people") - Can they trace bias to training data? (Look for: "We had more examples of X, so it learned X better") - Can they think about consequences? (Look for: "If we used this for decisions, group Y would be disadvantaged")

Extensions and Modifications

For more advanced students: - Discuss other sources of bias beyond training data (algorithm design, evaluation metrics, deployment context) - Have students propose statistical tests that could detect bias - Discuss fairness frameworks—does fairness mean equal accuracy, or something else?

For younger students: - Focus on concrete example (the activity itself) rather than abstract discussion - Use simpler language; skip technical terminology - Focus on the basic idea: "The algorithm learned better from examples it saw more"

For longer block schedules: - Have students conduct more extensive training and testing - Have different groups test each other's algorithms - Have students propose modifications and retrain

Why This Works

This lesson makes bias concrete. Students don't just hear that AI can be biased—they see it happen with their own algorithm. When they realize that their carefully trained algorithm is actually less accurate for some people, the insight sticks.

The hands-on nature means students are building intuition, not memorizing facts. When they leave class, they're more likely to question algorithmic fairness in the world around them.

And importantly, the lesson avoids cynicism. Bias isn't shown as a reason to reject AI. It's shown as a real challenge that requires attention. Students learn that bias can be reduced through careful data collection, evaluation, and design.

That's genuine AI literacy.