What AI Actually Is
A rigorous introduction to artificial intelligence: definitions, taxonomies, and the operational meaning of machine intelligence in modern computing.
Learning Objectives
- Define artificial intelligence using modern academic terminology
- Distinguish between AI, machine learning, and deep learning
- Understand what "intelligence" means operationally in computing systems
- Identify common misconceptions and their corrections
- Interact with a pattern-recognition decision boundary system
The Working Definition
Artificial Intelligence refers to computational systems designed to perform tasks that, when performed by humans, would be described as requiring reasoning, learning, perception, or decision-making under uncertainty.
Consider what happens when you recognize a friend's face in a crowd. Your visual system processes millions of light signals, extracts features (the shape of eyes, the curve of a jawline), compares them against stored memories, and produces a judgment—all in milliseconds, often despite poor lighting, unusual angles, or years since your last meeting.
AI systems attempt to replicate this kind of competent behavior: not by copying human biology, but by building computational structures that achieve similar functional outcomes.
The field divides roughly into two paradigms:
- Symbolic AI (1950s–1980s dominant): Systems that manipulate explicit symbols according to rules. Think of expert systems with if-then rules encoding human knowledge.
- Statistical/Connectionist AI (1990s–present dominant): Systems that learn patterns from data using mathematical optimization. This includes neural networks, which learn hierarchical representations by adjusting millions of numerical parameters through a process called gradient descent (covered in Module 1.4).
Modern AI is overwhelmingly statistical. When you hear about GPT, DALL-E, or AlphaFold, you're hearing about systems that learned their capabilities from data, not systems programmed with explicit rules.
A spam filter is AI: it classifies emails by learning patterns that distinguish spam from legitimate messages. A recommendation system is AI: it predicts your preferences by finding statistical structure in user behavior data. A self-driving car's perception system is AI: it segments camera images into objects by learning from millions of labeled examples.
The common thread: tasks that require generalization from experience—the ability to handle inputs the system has never seen before, based on patterns extracted from inputs it has seen.
The Taxonomy: AI → ML → DL
Any system exhibiting intelligent behavior
Systems that improve through experience
Learning hierarchical representations via neural networks
Machine Learning
Machine learning is the subset of AI where systems improve their performance on a task through exposure to data, without being explicitly programmed for every case. The canonical formulation comes from Tom Mitchell (1997):
"A computer program is said to learn from experience E with respect to some class of tasks T and performance measure P, if its performance at tasks in T, as measured by P, improves with experience E."
This is precise: learning means measurable improvement on defined tasks given more data. It excludes systems that merely store and retrieve information (databases) and systems whose behavior is entirely hand-coded (traditional software).
Deep Learning
Deep learning is a specific approach within machine learning that uses artificial neural networks with multiple layers (hence "deep") to learn hierarchical representations of data.
The key insight: rather than hand-engineering features (e.g., telling an image classifier to look for edges, then corners, then shapes), deep learning systems learn the features themselves. Early layers might learn simple patterns (edges), middle layers combine these into complex patterns (textures, object parts), and final layers assemble these into task-relevant concepts (faces, cats, tumors).
Deep learning dominates modern AI because it scales: given sufficient data and computation, deeper networks with more parameters consistently achieve better performance on complex tasks like image recognition, language understanding, and protein structure prediction.
Types of Machine Learning
Machine learning encompasses several fundamentally different learning paradigms, distinguished by how the system receives feedback and what it's trying to learn. Understanding these paradigms is essential vocabulary for modern AI.
Supervised Learning
Learning from labeled examples—input-output pairs where the correct answer is provided during training.
How it works: Given data points (X) with known labels (Y), learn a function f such that f(X) ≈ Y. Minimize prediction error on training data, then generalize to new inputs.
- Classification: Email spam detection (spam/not spam)
- Regression: House price prediction (continuous value)
- Image classification: Identifying objects in photos
This is the dominant paradigm in most of this course.
Unsupervised Learning
Finding structure in unlabeled data—discovering patterns without being told what to look for.
How it works: Given only inputs (X) with no labels, identify underlying structure: clusters, dimensions of variation, or density distributions.
- Clustering: Customer segmentation, grouping similar documents
- Dimensionality reduction: PCA, t-SNE for visualization
- Anomaly detection: Fraud detection, system monitoring
Reinforcement Learning
Learning through interaction and rewards—an agent learns by taking actions in an environment and receiving feedback.
How it works: Agent observes state, takes action, receives reward. Goal: learn a policy that maximizes cumulative reward over time. Handles the exploration-exploitation tradeoff.
- Game playing: AlphaGo, chess engines
- Robotics: Learning to walk, grasp objects
- RLHF: Fine-tuning ChatGPT with human feedback
Self-Supervised Learning
Creating supervision signal from the data itself—a modern paradigm powering large language models.
How it works: Create tasks where labels are derived from the input (e.g., predict masked words, predict next token). Enables learning from massive unlabeled datasets.
- Language models: GPT's next-word prediction
- Masked prediction: BERT filling in [MASK] tokens
- Contrastive learning: CLIP matching images with text
Key enabler of foundation models—covered in Modules 3.3-3.4.
Semi-Supervised Learning
Combining small labeled + large unlabeled datasets—leveraging the best of both worlds.
How it works: Use limited labeled data to guide learning, while extracting additional signal from abundant unlabeled data. Addresses the labeling bottleneck.
- Medical imaging (labels require expert annotation)
- Speech recognition with limited transcriptions
- Web content classification
Why This Taxonomy Matters
These aren't just academic categories—they represent fundamentally different learning settings with different algorithms, assumptions, and applications. Knowing whether your problem is supervised (you have labels), unsupervised (you need to discover structure), or reinforcement (you have rewards) is the first step in choosing the right approach.
What "Intelligence" Means Operationally
The word "intelligence" carries philosophical baggage. In AI, we sidestep debates about consciousness or understanding by focusing on operational definitions: intelligence is measured by what a system can do, not what it "is."
The Behavioral View
A system exhibits intelligence to the degree that it produces appropriate responses to novel situations. "Appropriate" is defined relative to some objective—winning a game, correctly classifying an image, generating coherent text.
This is why we can meaningfully say that a chess engine "plays intelligently" without claiming it thinks, feels, or understands chess the way a grandmaster does. The engine produces moves that achieve the objective (winning) in situations it has never encountered before. That capacity—generalization to new inputs—is the operational signature of intelligence in computing.
Intelligence vs. Mere Computation
| Characteristic | Traditional Program | Intelligent System |
|---|---|---|
| Handles novel inputs | Only if explicitly coded | Generalizes from training data |
| Adapts over time | Static until manually updated | Can improve with more experience |
| Knowledge source | Programmer's explicit rules | Patterns extracted from data |
| Failure mode | Crashes or wrong output | Graceful degradation, uncertainty estimates |
Common Misconceptions
"AI systems understand what they're doing"
This anthropomorphizes AI. Current systems are sophisticated pattern matchers that optimize objectives. A language model predicting the next word doesn't "understand" language in any human sense—it has learned statistical regularities that produce coherent outputs.
The accurate framing: AI systems exhibit functional competence without necessarily possessing understanding, awareness, or intentions. They solve problems through mathematical optimization, not comprehension.
"AI learns the way humans learn"
Current AI requires vastly more data than humans. A child learns to recognize cats from a handful of examples; an image classifier needs thousands. Human learning involves embodiment, curiosity, social context, and sleep—none of which apply to neural networks.
The accurate framing: Machine learning and human learning share the abstract principle of "improvement from experience" but differ radically in mechanism, efficiency, and requirements. Neural networks are inspired by brains but are not models of brains.
"More data always means better AI"
Data quality matters as much as quantity. Biased data produces biased models. Noisy labels degrade learning. And beyond a certain scale, returns diminish—doubling data might yield only marginal improvements.
The accurate framing: More data can improve performance, but only if the data is representative, correctly labeled, and relevant to the task. Data curation is often more valuable than data collection.
"AI will inevitably become sentient/conscious"
This assumes that scaling pattern matching leads to consciousness—an unsupported leap. We don't understand consciousness well enough to know if any computational process could produce it, let alone current architectures.
The accurate framing: Current AI systems are tools that process information according to trained parameters. Whether AI could ever be conscious is an open philosophical question, not an engineering inevitability.
Interactive Lab: Decision Boundaries
A fundamental task in AI is classification: given an input, assign it to one of several categories. The decision boundary is the surface that separates different classes in the input space.
In this lab, you'll see how a simple classifier learns to separate two classes of points. Add points, adjust the model complexity, and observe how the decision boundary changes.
What to Observe
- Linear boundaries (low complexity) can only separate data with a straight line. If your classes are interleaved, accuracy will suffer.
- High complexity models can fit intricate boundaries but may "overfit"—learning the noise in your specific points rather than the true pattern.
- The learning rate controls how aggressively the model updates. Too high and it oscillates; too low and it converges slowly.
- Notice how adding a single point in a critical region can dramatically change the boundary.
Check Your Understanding
According to the operational definition, what makes a system "intelligent"?
What distinguishes deep learning from other machine learning approaches?
Why is the claim "AI systems understand what they're doing" considered a misconception?
In the decision boundary lab, what happens when you increase model complexity for data that is linearly separable?
According to Tom Mitchell's definition, what three elements must be specified to define a machine learning problem?