1 / 33
Artificial Intelligence · Grade 10 · Chapter 3

What is AI? Foundations, Paradigms & Ethics

From symbolic rule engines and the Turing Test to statistical machine learning, neural networks, and algorithmic alignment.

Symbolic vs ML ANI vs AGI Supervised Learning Unsupervised Learning Reinforcement Learning
Chapter Roadmap

Learning Objectives

1. Foundations & Philosophy

Differentiate Narrow, General, and Super AI; evaluate the Turing Test and Searle's Chinese Room argument.

2. Paradigm Shift in Computing

Contrast traditional algorithmic programming ($Data + Rules$) with Machine Learning ($Data + Answers$).

3. The 3 Learning Paradigms

Master Supervised (Classification/Regression), Unsupervised (Clustering), and Reinforcement Learning (Agent/Reward).

4. Ethics & Alignment

Analyze data bias, explainability (black-box problem), hallucinations, and safety alignment.

A Tale of Two Breakthroughs

Deep Blue (1997) vs AlphaGo (2016)

Two historic milestones that define the evolution of artificial intelligence:

1997: IBM Deep Blue (Symbolic AI)

Defeated World Chess Champion Garry Kasparov using brute-force search algorithms (Minimax + Alpha-Beta pruning) evaluating 200 million positions/sec based on human-written chess rules.

2016: DeepMind AlphaGo (Deep RL)

Defeated Go Master Lee Sedol. The board has more states than atoms in the universe ($10^{170}$). Deep Blue's method would fail. AlphaGo learned intuition by playing millions of games against itself using deep neural networks!

Core Definitions

Defining Artificial Intelligence

In computer science (Russell & Norvig), AI is formally defined as the study of Rational Agents:

Perception

Receiving continuous sensory inputs from the environment (cameras, microphones, LIDAR, raw text).

Reasoning & Inference

Processing information against internal probabilistic models or knowledge graphs to determine the best path.

Autonomous Action

Taking actions in the environment that maximize the expected probability of achieving a defined goal.

The Capability Spectrum

The 3 Tiers of Artificial Intelligence

ANI

Narrow AI (ANI)

Today's AI: Specialized in a single domain. AlphaFold folds proteins, Midjourney draws images, Siri transcribes voice, but none can drive a car or write poetry on their own.

AGI

General AI (AGI)

Human-Level AI (Theoretical): An autonomous system that can learn, understand, adapt, and transfer cognitive skills across any intellectual task a human can perform.

ASI

Super AI (ASI)

Superintelligence (Hypothetical): An intelligence surpassing the collective cognitive capabilities of all human minds across science, creativity, and strategy.

Historical Context

The AI Timeline & The Deep Learning Boom

1956: Dartmouth Workshop

John McCarthy coins "Artificial Intelligence". Early optimism centered around symbolic logic and theorem provers.

1970s - 1990s: AI Winters

Funding collapsed twice when rule-based expert systems failed to handle real-world messy, ambiguous, noisy data.

2012 - Present: Deep Learning

ImageNet breakthrough (AlexNet). Three converging forces: Massive Internet Datasets + GPU Parallelism + Multi-layer Neural Networks.

Philosophy of Mind

Can Machines Truly "Understand"?

Alan Turing (1950) — The Imitation Game

  • Behaviorist View: If a machine communicates via text so convincingly that a human judge cannot distinguish it from a human, it exhibits intelligence.
  • Focuses on outward functional capability and performance.

John Searle (1980) — The Chinese Room

  • Syntactic vs Semantic: A person locked in a room using a rulebook to swap Chinese symbols produces perfect responses without understanding a single word.
  • Computers manipulate syntax ($0$ and $1$); they possess zero subjective semantic consciousness.
Architectural Paradigms

Symbolic AI vs Machine Learning

Symbolic / Expert Systems (Old AI)

Humans hand-craft thousands of deterministic IF-THEN rules. Fails completely when recognizing handwriting, faces, or natural spoken dialect.

if animal.has_stripes and animal.is_equine: return "Zebra" # Brittle! What if the photo is blurry or dark?

Machine Learning (Modern AI)

Instead of hardcoding rules, we feed an optimization algorithm 500,000 labeled animal images. The algorithm adjusts mathematical weights to discover features automatically!

Fundamental Paradigm Shift

Traditional Programming vs Machine Learning

Traditional Software Engineering

[ Data ] + [ Rules / Code ] ➔ [ Answers ]

Software developers write explicit deterministic logic for every edge case.

Machine Learning Engineering

[ Data ] + [ Answers / Labels ] ➔ [ Learned Model / Rules ]

The computer discovers the mathematical function mapping inputs to target outputs.

Core Enablers

The Machine Learning Trinity

Modern AI exists only at the intersection of three technical forces:

1. High-Quality Data

Billions of labeled images, text tokens, telemetry logs, and sensor arrays required to prevent statistical overfitting.

2. Optimization Algorithms

Gradient descent, backpropagation, transformer attention, and loss optimization math.

3. Massive Compute (GPUs)

Clusters of tensor processing units performing billions of floating-point matrix multiplications per second.

Taxonomy of Machine Learning

The 3 Major Learning Paradigms

1

Supervised

Learning with a teacher. Training data contains both inputs ($X$) and ground-truth labels ($y$).

2

Unsupervised

Learning without labels. Discovering hidden clusters, anomalies, and structures in raw data.

3

Reinforcement

Learning through interaction. An autonomous agent maximizes numerical rewards via trial and error.

Supervised Learning Subtypes

Classification vs Regression

Classification (Discrete Labels)

Predicting categorical classes or binary states:

  • Spam Detection: Is this email Spam or Not Spam?
  • Medical Imaging: Is tumor Malignant or Benign?
  • Vision: Is image a Cat, Dog, or Horse?

Regression (Continuous Values)

Predicting a real-valued continuous number:

  • Real Estate: Predicting house price based on sqft & location ($452,300).
  • Meteorology: Predicting tomorrow's high temperature (28.4°C).
  • Finance: Forecasting stock price movements over time.
Unsupervised Learning

Pattern Discovery in Unlabeled Data

The algorithm receives raw vectors without any human labels and finds latent mathematical geometry:

Clustering (K-Means)

Grouping customers into behavioral segments based on purchasing history without predefined categories.

Anomaly Detection

Flagging credit card transactions or server network traffic that deviates from normal statistical distributions.

Dimensionality Reduction

Compressing 10,000 gene features into 2D plots (PCA / t-SNE) while preserving relational clusters.

Reinforcement Learning

The Reinforcement Learning Loop

How robots and game agents learn to navigate complex dynamic environments:

S

State ($S_t$)

Agent observes current board position or robot sensor readings.

A

Action ($A_t$)

Agent selects an action based on its policy network $\pi(a|s)$.

E

Environment

Physics world advances state to $S_{t+1}$ and evaluates outcome.

R

Reward ($R_{t+1}$)

Agent receives $+10$ (goal reached) or $-100$ (crashed); updates weights!

Decision Strategy

The Exploration vs Exploitation Tradeoff

A core mathematical challenge in Reinforcement Learning:

Exploitation

Choose the best known action that previously yielded the highest reward. (e.g. Always ordering your favorite dish at a restaurant).

Risk: May get stuck in a suboptimal local maximum.

Exploration

Try unknown actions to discover potentially superior strategies. (e.g. Trying a random new dish on the menu).

Benefit: Discovers novel superhuman strategies (like AlphaGo's Move 37).
Sensory AI

Major AI Perception Modalities

Computer Vision (CV)

Object detection (YOLO), semantic segmentation, facial recognition, autonomous driving visual perception.

Natural Language Processing

Machine translation, sentiment analysis, named entity recognition, and LLM text generation.

Audio & Speech AI

Automatic Speech Recognition (ASR: Whisper), speech synthesis (TTS), audio acoustic event detection.

Hierarchical Feature Learning

Why Deep Neural Networks Dominate

Classical algorithms required human engineers to hand-design feature extractors. Deep neural networks learn features in hierarchical layers:

L1

Early Layers

Detect low-level edges, diagonal gradients, and color blobs.

L2

Middle Layers

Combine edges into textures, shapes, noses, wheels, and eyes.

L3

Deep Layers

Synthesize complete object concepts: human faces, cars, stop signs.

Debunking Sci-Fi Myths

Misconception: "AI is Conscious and Has Feelings"

The Hollywood Illusion

"ChatGPT sounds friendly and empathetic, so it must have emotions, consciousness, and personal desires."

The Mathematical Reality

Large Language Models are next-token probability distribution calculators ($P(w_n | w_1, w_2, \dots, w_{n-1})$). They simulate human conversational style through statistical correlation of trillions of words, with zero emotional awareness.

AI Safety & Ethics

Algorithmic Bias: Garbage In, Garbage Out

An AI model is only as fair as the historical data used to train it.

Case Study: Automated Resume Screening Bias

In 2018, a major tech firm trained an AI tool to score job resumes using 10 years of historical hiring data. Because past tech hires were predominantly male, the AI learned to penalize resumes containing the word "Women's" (e.g. "Women's Chess Captain").

Lesson: AI models don't eliminate human prejudice—they mathematically amplify historical biases unless actively audited and corrected!
Explainable AI (XAI)

The "Black Box" Problem

A deep neural network may contain over 500 billion parameters. Even the researchers who trained it cannot explain precisely why it made a specific prediction.

High-Stakes Hazards

If an AI model denies someone a mortgage, rejects a cancer patient's treatment, or flags a defendant as "high flight risk", the lack of explainability violates due process and legal compliance.

Explainable AI Solutions

Researchers develop attention heatmaps (Grad-CAM) and feature attribution tools (SHAP / LIME) to illuminate which input features drove the model's decision.

Interactive Smartboard Activity

Match the Real-World Problem to the ML Paradigm

SMARTBOARD VOTING

1. Predicting house selling prices from square footage & bedrooms  →  [Supervised: Regression]
2. Teaching a simulated quadruped robot to walk across rough terrain  →  [Reinforcement Learning]
3. Grouping 100,000 news articles into topics without any preset tags  →  [Unsupervised: Clustering]
4. Flagging whether a biopsy image contains skin cancer  →  [Supervised: Classification]

Ethics Discussion

Think-Pair-Share: The Self-Driving Car Dilemma

An autonomous car with failing brakes must choose between swerving into a concrete wall (harming the passenger) or staying on course (harming 3 pedestrians crossing illegally).

The Algorithmic Dilemma

Unlike a human driver making a split-second panic reaction, a self-driving car's decision was pre-programmed by software engineers months in advance.

Discussion Prompt (3 Mins)

Who should be held legally and morally accountable when an AI agent causes harm: the programmer, the car manufacturer, the passenger, or the government regulators?

Guided Engineering Practice

Formulating an AI Solution

Your school wants to build an AI system to predict whether students need additional tutoring in math.

1

Input Features ($X$)

Quiz scores, homework completion rate, class attendance, time spent on digital portal.

2

Target Label ($y$)

Binary: 1 (Needs intervention) or 0 (On track).

3

Paradigm

Supervised Binary Classification (e.g. Logistic Regression or Random Forest).

AI Auditing Challenge

Spot the Flaw in this AI Training Plan

Proposed Project: "Smart Wildlife Night Camera"

"We collected 10,000 photos of wolves and 10,000 photos of domestic dogs from Google Images. We trained a deep convolutional network and achieved 99.8% test accuracy! However, when deployed in the forest, the model misclassifies every single dog as a wolf whenever it snows."

Analyze: Why did this model fail in the real world? What spurious correlation did it learn?
Spurious Correlations Explained

The "Snowy Wolf" Classic ML Trap

The Hidden Flaw

In the training dataset, almost all wolf photos had snow in the background, while domestic dog photos were taken on green grass or living rooms. The neural network never learned what a wolf looks like—it simply learned to detect white snow pixels!

The Computer Science Lesson

Neural networks are opportunistic shortcut learners. If there is a simple spurious shortcut (snow = wolf), the model will exploit it instead of learning the true underlying object features.

Knowledge Check 1 / 4

Which statement correctly characterizes Artificial Narrow Intelligence (ANI)?

AIt possesses human-level consciousness and general reasoning across all tasks
BIt is specialized to excel at a single specific task or domain
CIt has surpassed all human intelligence and can self-replicate
DIt relies exclusively on mechanical gears without computers
Click to reveal answer
Explanation: All existing modern AI (Siri, ChatGPT, AlphaFold, Tesla Autopilot) is Artificial Narrow Intelligence (ANI), excelling at specific bounded domains.
Knowledge Check 2 / 4

How does Machine Learning differ from Traditional Programming?

AML does not use any mathematical calculations
BTraditional programming feeds rules to get answers; ML feeds data & answers to discover the rules
CML only works on quantum supercomputers
DTraditional programming cannot run in Python
Click to reveal answer
Explanation: In ML, algorithms learn mathematical mappings (rules/weights) directly from paired inputs and ground-truth answers.
Knowledge Check 3 / 4

Which learning paradigm involves an autonomous agent receiving numerical rewards?

ASupervised Classification
BUnsupervised Clustering
CReinforcement Learning
DSymbolic Expert Systems
Click to reveal answer
Explanation: Reinforcement Learning trains agents via environmental feedback (state transitions and scalar rewards/penalties).
Knowledge Check 4 / 4

What was the core conclusion of John Searle's "Chinese Room" thought experiment?

AComputers can easily understand human emotional subtleties
BManipulating symbols/syntax does not constitute genuine semantic understanding or consciousness
CThe Turing Test is mathematically impossible to pass
DNeural networks are completely useless for translation
Click to reveal answer
Explanation: Searle argued that purely syntactic symbol manipulation (as done by digital computers) does not produce true semantic comprehension or intentionality.
Future Policy

AI Alignment & Responsible Deployment

Value Alignment

Ensuring AI objective functions align with human safety, ethics, and universal human rights.

Red Teaming & Auditing

Stress-testing models before public release to find vulnerabilities, jailbreaks, and harmful outputs.

Watermarking & Provenance

Embedding cryptographic signatures into AI-generated media to curb deepfake disinformation.

Chapter 3 Review

AI Foundations Master Summary

1. Intelligence Spectrum

ANI (domain-specific tools) ➔ AGI (human-level generalists) ➔ ASI (superintelligence).

2. The 3 Learning Paradigms

Supervised (labeled data), Unsupervised (latent structure), Reinforcement (reward optimization).

3. Deep Learning Power

Multi-layer neural networks extract hierarchical feature representations automatically from raw data.

4. Ethical Imperative

Addressing bias, explainability, safety alignment, and responsible deployment is vital for society.

Wrap-Up & Reflection

Exit Ticket: The AI Paradigm Matrix

Pick an industry you care about (e.g. Video Games, Medicine, Music, Space Exploration). Propose one application for each paradigm:

Supervised Learning

What input $X$ and label $y$ would you predict?

Unsupervised Learning

What hidden patterns or clusters would you discover?

Reinforcement Learning

What would the agent do and what is the reward function?

Coming Up Next

Looking Ahead: Real-World AI Applications

How do the algorithms we learned today power TikTok's recommender engine, iPhone Face ID, Tesla Autopilot, and ChatGPT?

Next in Chapter 4:

We will tear down real-world production AI systems—from collaborative filtering and computer vision pipelines to next-token prediction and diffusion models!