Machine learning, AI, and how software learns — the full module map
Before we start — three quick questions. Have a think first, then reveal.
Loads of options: Face ID (face recognition), Siri or Google Assistant (natural language processing), the keyboard autocomplete (next-word prediction), Instagram/TikTok feed ranking (recommendation system), Google Photos auto-tagging faces or scenes, spam filtering in your inbox. You're using AI constantly without thinking about it.
Rule-based: someone writes explicit if/else logic. "If email contains 'Nigerian prince', mark as spam." You need to anticipate every case upfront. It breaks the moment spam evolves.
Learning system (ML): you feed it thousands of examples of spam and not-spam, and it figures out the rules itself. It adapts as new spam patterns emerge — you just give it more labeled data.
Key difference: who writes the rules — a programmer, or the data.
Yes, many times — this is one of the most important effects of this module. Examples:
The pattern: automation destroys some jobs and creates others (often requiring different skills). Whether that's net-positive is a values question, not just a technical one — and it's in your HSC syllabus.
Automation in software isn't one thing — it's a spectrum from "scripts that do stuff" all the way to "systems that teach themselves." The HSC curriculum covers three specific paradigms:
e.g. a Python script that renames 500 files
Bots that mimic human clicks and keystrokes
Automated workflows across whole systems
Systems that learn from data
Every time you push to GitHub, a pipeline can automatically run your tests, check for vulnerabilities, and deploy if everything passes — no human needed.
GitHub Actions is exactly this. Your Volcanic Pantheon repo could use it.
RPA bots log into systems, copy data, fill out forms, and send emails — exactly like a human would, but 24/7 and without mistakes. Used heavily in finance and healthcare for repetitive admin.
Example: A bot that extracts invoices from email and enters them into accounting software.
Automating entire multi-step business processes — not just one action, but chains of them across different systems.
Example: Customer submits support ticket → auto-classified by topic → routed to correct team → customer receives confirmation email. All automatic.
HSC note: The syllabus groups these under "software automation" because they all replace or augment human effort with software. ML/AI is the most advanced tier — and the main focus of this module.
These terms get used interchangeably, which makes them confusing. In the HSC — and in professional settings — they mean different things:
| Term | Definition | Who writes the rules? |
|---|---|---|
| AI | Any system that simulates intelligent behaviour — including ones with hand-coded rules | A human programmer |
| ML | A subset of AI where the system learns patterns from data rather than following explicit rules | The data |
Imagine you want a thermostat that activates at exactly 18°C. A programmer writes: if temperature < 18: turn_on() — that's AI (rule-based). Now imagine a thermostat that watches your behaviour for a month and figures out when you're comfortable without you specifying any temperature. It learned from data — that's ML.
Deep Blue (1997) — beat Garry Kasparov using a rule-based evaluation function hand-written by grandmasters. Definitely AI, not ML.
AlphaZero (2017) — given only the rules of chess, it played against itself millions of times and learned the strategies on its own. Beat every other chess engine in under 24 hours. That's ML.
For each scenario, decide: is this rule-based AI, or machine learning? Then reveal.
Rule-based AI. A human wrote those keyword rules explicitly. The system doesn't learn — it just checks the list. This is how early spam filters worked, and spammers defeated them easily by slightly changing their wording.
Machine Learning. Google trained it on billions of labeled emails (spam/not-spam). It identifies patterns — not just keywords, but writing style, sender history, structure, time sent — that no human wrote explicitly. It also improves as you mark things as spam.
Neither — this isn't even AI. It's just a timer. AI implies some form of decision-making in response to inputs. However...
A smart traffic light that uses camera feeds to estimate queue lengths and adjusts timing accordingly? That starts to look like AI. One that learns peak traffic patterns over weeks and pre-adjusts? That's ML.
Machine Learning. Netflix trains models on what millions of users watched, rated, rewatched, or abandoned. Your recommendations are generated by finding users similar to you and surfacing what they liked. No human wrote "if they watched Breaking Bad, suggest Better Call Saul" — the model learned that correlation from data.
Rule-based AI. If it pattern-matches the question against a list of triggers and returns stored text, that's a scripted/rules-based chatbot — no learning involved.
ChatGPT, on the other hand, is ML — it was trained on enormous amounts of text and generates responses by predicting likely next tokens. Very different architecture.
ML systems are categorised by how they learn — specifically, whether the training data has labels (correct answers) attached to it.
| Type | How It Learns | Example You Know | Key Use Cases |
|---|---|---|---|
| Supervised | Labeled data → learn to predict the label on new examples | Spam filter: each training email is labeled "spam" or "not spam" | Classification, prediction, image recognition |
| Unsupervised | No labels — find hidden structure or groupings in raw data | Spotify grouping songs by acoustic similarity without being told genre | Clustering, anomaly detection, compression |
| Semi-supervised | A small amount of labeled data + a large unlabeled dataset | Google Photos — you label a few photos of your friend, it finds the rest | Face recognition, medical imaging (labeling is expensive) |
| Reinforcement | Agent takes actions, receives rewards or penalties — maximises score over time | Chess/Go AI (AlphaZero), video game bots, robotics | Games, autonomous vehicles, resource scheduling |
Think about how you got good at a game. You didn't read a manual of every possible move. You played, made mistakes, saw what worked, and gradually improved your strategy. Reinforcement learning works the same way — the agent (the software) takes actions in an environment, receives a reward signal (score, win/loss), and over thousands of iterations learns which actions lead to better outcomes.
This is how DeepMind's AlphaGo defeated the world Go champion in 2016. Go has more possible game states than atoms in the observable universe — you can't hand-code good moves. It had to learn them.
Supervised learning. You'd need a training dataset of images already labeled with character names. The model learns to associate visual features (hair style, colour, clothing, accessories) with the correct character label. Then when a new image is uploaded, it predicts the most likely character.
This is exactly how Google's image search works — trained on billions of labeled images.
You don't need to implement these from scratch — but you do need to understand what each one does and when you'd use it. Think of this as a toolkit: each algorithm has a strength.
A flowchart of yes/no questions. At each node, the model asks a question about a feature. You follow the branches until you reach a leaf — a prediction.
Intuitive and explainable. You can literally print the tree and read it. Great for classification.
Fits a straight line through data points to predict a number. Given X (input), estimate Y (output). The model learns the slope and intercept from your training data.
Output is continuous — a price, a duration, a temperature. Not a category.
Despite the name, it's used for classification. It outputs a probability between 0 and 1, then you pick a threshold (e.g. >0.5 = yes). The output is an S-curve, not a line.
Best when you need a confidence score, not just a label.
Classify a new point by looking at the K closest points in your training data and taking a majority vote. No training phase — it just stores the dataset and reasons at prediction time.
Simple but can be slow on large datasets.
The algorithm powering most modern AI — image recognition, language models, voice assistants. Loosely inspired by biological neurons in the brain.
A neural network has three types of layers:
Each connection between nodes has a weight. During training, the network adjusts these weights to minimise its error — this process is called backpropagation.
Why are neural networks so powerful? They can learn arbitrary non-linear patterns from data — patterns that no programmer could write rules for. But they're also a "black box" — it's very hard to explain why they made a particular decision. This has serious ethical implications.
For each problem, which algorithm would you reach for first? Reveal to see the reasoning.
Linear regression (or polynomial regression if the relationship is non-linear). You're predicting a continuous number — price — from numeric and categorical inputs. This is the classic regression problem.
Neural network — specifically a convolutional neural network (CNN). Images have too many dimensions for simpler algorithms to handle well. Neural networks excel at finding spatial patterns in pixel data. This is exactly what Google Photos and iPhone's scene detection do.
Logistic regression or decision tree — you want a clear yes/no classification (approve/deny), and you also need to be able to explain the decision (regulatory requirement). A decision tree is especially good here because you can show a customer exactly which rule triggered the denial. Neural networks would be poor — they're a black box, which is legally problematic in lending.
Unsupervised learning — clustering algorithm (e.g. K-Means). You have no labels — you don't know in advance what the groups are. The algorithm groups customers by similarity in their purchase data. A retailer uses this to build customer segments for targeted marketing.
Note: KNN is supervised (uses labeled training data). Clustering algorithms like K-Means are unsupervised. Don't confuse them — K in KNN and K in K-Means are different things.
The HSC syllabus specifically lists these ML application areas — all of which you already interact with daily:
| Application Area | What It Does | Where You've Seen It |
|---|---|---|
| Data Analysis & Forecasting | Find patterns in large datasets; predict future values | Spotify Wrapped, stock market prediction, weather forecasting |
| Virtual Assistants | Understand natural language, generate responses | Siri, Alexa, ChatGPT, Google Assistant |
| Image Recognition | Identify objects, faces, or scenes in photos/video | Face ID on your phone, Google Photos auto-tagging, medical scans |
| Recommendation Systems | Predict what you'll like based on your history and similar users | YouTube autoplay, Netflix, TikTok For You page |
If you ever add comments, a text classification model could auto-flag toxic or off-topic content before it appears.
ML type: Supervised (trained on labeled "toxic/ok" text)
A recommendation system that suggests related characters or pages based on what similar visitors browsed.
Algorithm: KNN or collaborative filtering
Upload an image → model identifies the character and auto-fills the tag. Removes manual curation effort.
Algorithm: Convolutional neural network
The HSC devotes a full topic to the societal and ethical implications of ML. We'll go deep in Lesson 20 — but here's a preview of why it matters:
ML models don't have opinions — they learn whatever patterns exist in their training data. If that data is biased, the model inherits and amplifies the bias.
A 2019 MIT study found that several commercial facial recognition systems misidentified darker-skinned women at rates up to 34% higher than lighter-skinned men. Why? The training datasets were mostly white, male faces. The model learned those patterns better.
These same systems are used by some law enforcement agencies to match faces from security footage.
Automation eliminates some roles and creates others. The net effect depends on how fast new roles emerge compared to how fast old ones disappear — and who has access to retraining. Historically, the workers displaced first are those in routine, repeatable jobs. The ones created are often technical and highly skilled.
One of the most positive applications: robots and automated systems doing dangerous work — mining, demolition, handling toxic materials, deep-sea repair. The HSC wants you to acknowledge both sides.
Is it ethical to use ML to screen job applications — automatically filtering out candidates before a human sees their CV? What could go wrong?
Arguments for: Reduces unconscious human bias; consistent criteria applied to all; faster, scalable at volume.
Arguments against:
Three lessons cover the full Software Automation module. Here's where we are:
Automation spectrum, AI vs ML, types of learning, key algorithms, applications, ethics teaser
Building a regression model in Python; decision tree implementation; neural network predictions
Full societal impact analysis: bias, employment, disability, economy, psychological effects; HSC exam questions
The difference between AI and ML — and why it matters for exam answers
The four types of ML learning and a real example of each
What a decision tree, linear regression, logistic regression, KNN, and neural network each do
Three concrete automation paradigms: DevOps/CI-CD, RPA, BPA
Why bias in training data is a real and serious problem
Task 1 — Spot ML in the wild
Find a feature on any website or app that you think uses ML. Write two sentences:
1. What type of ML do you think it uses (supervised / unsupervised / reinforcement)?
2. Why do you think that — what's your evidence?
Task 2 — Design for Volcanic Pantheon
Pick one of the three ML feature ideas from Section 5 (comment moderation, character recommendations, or fanart auto-tagging) and write a short paragraph covering:
1. What type of ML it uses
2. What training data it would need (be specific — where would you get it?)
3. One ethical risk with using it on your site
There's no single right answer for either task — the reasoning is what matters. We'll discuss at the start of next lesson.