AI & Digital Ethics Grade 8 Computer Vision How Machines See

Grade 8 – Computer Science (AI & Digital Ethics)
Topic: Computer Vision: How Machines See

1. The Driving Question

If a self-driving car "sees" a stop sign, does it actually understand what a stop sign means—or is it just guessing based on pixels? And if it guesses wrong, who’s responsible: the programmer, the AI, or the person who painted the sign? How do we teach a machine to see the world like we do, when all it really knows is numbers?

2. The Core Idea – Built, Not Listed

Imagine you’re playing a game of "I Spy" with a robot. You say, "I spy something red, octagonal, and covered in white letters," and the robot scans the street. But instead of recognizing a stop sign, it might see a red kite, a fire hydrant, or even a blurry smudge. That’s because computers don’t "see" like humans—they break images into tiny squares called pixels, each with a number representing color (like #FF0000 for red). A computer vision system, like the one in a phone’s camera or a security drone, uses algorithms (step-by-step math rules) to analyze these numbers and make guesses: "This cluster of red pixels is 85% likely to be a stop sign."

But here’s the catch: the machine doesn’t know what a stop sign is. It doesn’t understand that stopping prevents crashes or that the sign’s shape is a law. It only knows that in 10,000 training photos, red octagons with white letters were labeled "stop sign." If someone sticks a sticker on the sign or the lighting changes, the AI might fail—because it’s not seeing, it’s pattern-matching. And that raises a big question: If an AI’s "vision" is just statistical guesswork, can we trust it to make life-or-death decisions, like in a self-driving car?

Key Vocabulary:
- Pixel – The smallest unit of a digital image, like a single tile in a mosaic. Each pixel is a tiny square with a color value (e.g., a pixel in a stop sign might be #FF0000 for red).
Example: If you zoom in on a photo of your dog, you’ll see it’s made of thousands of colored squares—those are pixels.

• Algorithm – A set of step-by-step instructions a computer follows to solve a problem. In computer vision, algorithms might count edges, compare colors, or measure shapes.
  Example: A facial recognition algorithm might measure the distance between your eyes and the shape of your jawline to guess who you are.

• Training Data – The thousands (or millions) of labeled examples an AI learns from. If the data is biased (e.g., mostly photos of stop signs in daylight), the AI might fail in new situations.
  Example: If an AI is trained only on photos of cats and dogs, it might mistake a raccoon for a cat—because it’s never seen one before.

• Bias (in AI) – When an AI’s decisions unfairly favor or harm certain groups because of flaws in its training data or design.
  Example: Some early facial recognition systems worked poorly on darker skin tones because they were trained mostly on photos of light-skinned people.

(Note for high school/college: In advanced AI, "seeing" isn’t just about pixels—it’s about feature extraction (identifying edges, textures, or even emotions) and neural networks (layers of algorithms that mimic the human brain). Bias isn’t just a data problem; it’s a systemic issue tied to who builds the AI and what problems they prioritize.)

3. Assessment Translation

How this appears on state tests (e.g., NGSS-aligned computer science assessments):
- Multiple Choice: Questions test understanding of how computer vision works (e.g., "Which of these is NOT a step in how a self-driving car ‘sees’ a pedestrian?") with distractors like: - "The car uses X-ray vision to see through clothes." (Misconception: AI has superhuman abilities.) - "The car remembers every pedestrian it’s ever seen." (Misconception: AI has perfect memory.) - "The car compares the image to a database of shapes." (Correct, but incomplete—distractor might omit the role of training data.)

• Short Answer: "Explain why a computer vision system might fail to recognize a stop sign if it’s covered in snow. Use the terms ‘training data’ and ‘algorithm’ in your answer."
• Proficient response: "The algorithm was probably trained on photos of stop signs in clear weather, so it learned to recognize red octagons with white letters. If snow covers the sign, the colors and shapes change, and the algorithm might not have enough examples of snowy stop signs in its training data to make the right guess."

• Developing response: "The computer can’t see it because it’s snowy." (Missing key terms and explanation.)

• Evidence-Based Writing: "A city wants to use AI-powered cameras to catch drivers running red lights. Some residents argue this is unfair because the AI might make mistakes. Write a paragraph arguing for or against using the system, using evidence about how computer vision works."

• Proficient response: "While AI cameras could reduce accidents, they shouldn’t be the only way to catch red-light runners because computer vision isn’t perfect. For example, if a camera is trained mostly on daytime photos, it might fail at night or in bad weather, leading to false tickets. Also, if the training data is biased (e.g., mostly photos of certain car models), it could unfairly target some drivers. Until AI can explain its decisions, humans should review the footage."

SAT/ACT Note (for high school):
Computer vision rarely appears directly on the SAT/ACT, but the logic behind it (e.g., interpreting data, understanding bias) shows up in Science and Reading sections. For example: - A passage might describe an AI system and ask, "Which statement best explains why the system’s accuracy dropped in low light?" (Answer: The training data lacked examples of dimly lit images.)

4. Mistake Taxonomy

Mistake 1: Overestimating AI "Understanding"
- Prompt: "True or False: A self-driving car’s computer vision system understands that a stop sign means ‘stop to avoid a crash.’ Explain your answer." - Common Wrong Response: "True, because the car stops when it sees the sign, so it must understand what it means." - Why It Loses Credit: The response confuses behavior (stopping) with understanding (knowing why stopping matters). The AI doesn’t grasp the concept of safety—it just follows a rule: "If pixels = stop sign, then brake." - Correct Approach: "False. The car’s system recognizes the pattern of a stop sign (red octagon, white letters) but doesn’t understand the reason for stopping. It’s like a parrot repeating ‘stop’ without knowing what the word means. The AI only knows that in its training data, stop signs were labeled as things to stop for."

Mistake 2: Ignoring Bias in Training Data
- Prompt: "A facial recognition AI works well on light-skinned faces but poorly on dark-skinned faces. What’s the most likely reason?" - Common Wrong Response: "The AI is racist." (This is a moral judgment, not a technical explanation.) - Why It Loses Credit: The answer doesn’t explain how bias enters the system. Assessments want students to connect bias to training data or algorithm design.
- Correct Approach: "The AI was probably trained mostly on photos of light-skinned faces, so it learned to recognize features common in those images (e.g., lighter skin tones, certain nose shapes). When it sees a dark-skinned face, it’s like trying to find a book in a library where most books are in a different language—it doesn’t have enough examples to make an accurate guess."

Mistake 3: Misapplying the Term "Algorithm"
- Prompt: "Describe one way an algorithm helps a computer vision system identify a cat in a photo." - Common Wrong Response: "The algorithm looks at the cat and knows it’s a cat." (This treats the algorithm like a human.) - Why It Loses Credit: The response doesn’t explain how the algorithm works—it just restates the goal.
- Correct Approach: "The algorithm might count the number of edges (lines) in the photo to find shapes, then compare those shapes to a database of cat features (e.g., pointy ears, whiskers). For example, it could measure the distance between the eyes and the shape of the ears to guess if the animal is a cat, dog, or raccoon."

5. Connection Layer

1. Within Computer Science: Computer vision → Machine learning — Computer vision relies on machine learning (a type of AI that learns from data). Understanding how an AI "sees" helps you grasp why machine learning models need diverse training data—if the data is limited, the AI’s "vision" will be too.

2. Across Subjects: Computer vision → Biology (human vision) — The way AI processes images is inspired by how our brains work! For example, both human eyes and computer vision systems detect edges first (like the outline of a stop sign), then fill in details. Studying one helps you understand the other.

3. Outside School: Computer vision → Social media filters — When you use a Snapchat filter that turns you into a dog or adds sunglasses, that’s computer vision! The app scans your face, identifies key features (eyes, nose, mouth), and overlays the filter. Next time you use one, notice how it struggles if you tilt your head or the lighting changes—just like a self-driving car in the rain.

6. The Stretch Question

"If a computer vision system is trained only on photos of stop signs in the U.S., could it recognize a stop sign in Japan, where the signs say ‘止まれ’ instead of ‘STOP’? Why or why not—and what would it take to make the AI work globally?"

Pointer Toward the Answer:
The AI might recognize the Japanese stop sign if it’s trained to look for shapes (octagons) rather than text. But if it’s only seen U.S. signs, it could fail because the colors, fonts, or even the background (e.g., Japanese signs might have different poles or landscapes) might throw it off. To make it work globally, you’d need: 1. Diverse training data (photos of stop signs from every country).
2. Flexible algorithms (ones that prioritize shape over text or color).
3. Testing in real-world conditions (e.g., signs covered in snow, graffiti, or different lighting).

This isn’t just a technical problem—it’s a cultural one. Who decides what a "stop sign" looks like to an AI? And what happens if a country’s signs don’t fit the AI’s definition?