INNOVAREModule 2 · Applied

Case 1: The Learning Machine

A spam filter, a Spotify playlist, and ChatGPT’s responses were all built differently. The difference comes down to one question: where does the training signal come from?

July 2026 · Case 1 of 6
As you read — hold this question

If a model is trained to predict text rather than to be correct, what should you always verify before trusting its output?

3 ways
machines learn — and the choice between them is not a technical preference. It determines what data you need, what you can train on, and what the machine can and cannot do. Most AI failures trace back to picking the wrong one.

Every machine learning system answers the same question: what counts as correct? In supervised learning, a human labels the answer. In unsupervised learning, there is no label — the machine finds structure that wasn’t defined in advance. In reinforcement learning, the machine tries things and receives rewards or penalties. These are not interchangeable. The training signal shapes everything the model can do — and everything it cannot.

The three paradigms
Quiz: Learning Types

What distinguishes them — not by name, but by mechanism

Paradigm Training signal Data requirement What it learns Where it fails
Supervised Human-labelled examples (input → correct output) Large labelled dataset — expensive and slow to create To map inputs to known output categories or values When labels are wrong, biased, or missing. Can only categorise what it saw in training.
Unsupervised None — the machine defines its own signal by finding patterns Large unlabelled dataset — much easier to collect Hidden structure: clusters, patterns, anomalies, relationships No way to verify if the patterns it found are meaningful. Human interpretation still required.
Reinforcement Reward function — defined by the designer, not labelled data An environment to act in, plus a reward signal A policy: which action maximises cumulative reward over time Reward hacking — optimising the signal rather than the intent. Designing the right reward is hard.
Applied — three paradigms, three real systems

What each paradigm looks like when deployed at scale

Supervised
AlphaFold

DeepMind’s AlphaFold (2020) solved the 50-year protein folding problem. Trained on a database of ~170,000 proteins with experimentally determined structures — labelled by scientists over decades. Given an amino acid sequence, it predicts 3D structure with near-atomic accuracy.

200M+ protein structures predicted and released free. New drug targets discovered in months, not years. The labels existed — AlphaFold learned to generalise from them.
Unsupervised
Spotify Discover Weekly

Spotify’s Discover Weekly uses collaborative filtering — an unsupervised approach that groups users by listening behaviour without any human saying “these users are similar.” The system finds latent clusters in 600M+ users’ play histories and maps tracks to those clusters.

Nobody labelled “jazz fan” or “workout playlist person.” The machine found those structures itself — and built a product that 70M+ users open weekly.
Reinforcement + RLHF
ChatGPT

GPT-4’s base model was self-supervised (trained to predict the next token). RLHF — Reinforcement Learning from Human Feedback — added the alignment layer: human raters score outputs, a reward model learns those preferences, and RL fine-tunes the base model toward responses humans rate higher.

A model that doesn’t just predict likely text — it predicts text that humans have rated as helpful, honest, and harmless. The training signal is human preference, not labels.
The fourth paradigm — and the most important one right now

Self-supervised learning: training without any human labels at all

Every frontier LLM — GPT-4, Claude, Gemini — begins with self-supervised pre-training. The model is given a sequence of text with words masked out and asked to predict them. The labels come from the text itself. This scales to the entire internet.

Why it scales
No human annotation required — the internet is the training set
Learns syntax, semantics, facts, and reasoning patterns simultaneously
Enables emergence — capabilities not explicitly trained appear at scale
What it doesn’t give you
Alignment — predicting internet text ≠ being helpful or safe
Accuracy — the model learns what text looks like, not what’s true
This is why RLHF is required as the second stage
The Principle — Training signal is everything
A model can only do what its training signal rewarded. A model trained to predict text predicts text — it doesn’t “know” if it’s correct. A model trained on biased labels learns the bias. A model with a poorly designed reward function will optimise the reward rather than the intent. Understanding which paradigm built a system tells you exactly what it can and cannot do — and where to look when it fails.
Take this away

A model can only do what its training signal rewarded. Understanding which paradigm built a system tells you exactly what it can and cannot do — and where to look when it fails.

Quick recall — without looking back

Test yourself on this case

Question 1 of 3

What is the difference between supervised, unsupervised, and reinforcement learning — in terms of training signal?

Supervised: a human labels the correct answer (input → output). Unsupervised: no label — the machine finds structure itself. Reinforcement: a reward function the designer defines — the machine tries actions and receives rewards or penalties. The key difference is where the “correct” signal comes from.
Question 2 of 3

Self-supervised learning is described as “the most important paradigm right now.” Why — and what does it not give you?

Self-supervised learning trains without any human labels — the labels come from the data itself (e.g. predict masked words). This scales to the entire internet and enables emergence — capabilities not explicitly trained appear at scale. What it doesn’t give: alignment (predicting internet text ≠ being helpful or safe) or accuracy (the model learns what text looks like, not what’s true). This is why RLHF is required as the second stage.
Question 3 of 3

AlphaFold, Spotify Discover Weekly, and ChatGPT each use a different paradigm. Match them — and explain why each needed that paradigm.

AlphaFold → Supervised: 170,000 labelled protein structures existed, so supervised learning was correct — it learns to generalise from those labels. Spotify → Unsupervised/collaborative filtering: no human labelled “jazz fan” — the machine found latent clusters in 600M+ users’ play histories. ChatGPT → Self-supervised pre-training + RLHF: internet text at scale for pre-training; human preference ratings provide the alignment layer.

Sources

Mitchell (1997)
Mitchell, T. (1997). Machine Learning. McGraw-Hill.
AlphaFold
Jumper, J. et al. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596, 583–589.
Spotify
Koren, Y., Bell, R., & Volinsky, C. (2009). Matrix factorization techniques for recommender systems. Computer, 42(8), 30–37.
RLHF
Ouyang, L. et al. (2022). Training language models to follow instructions with human feedback (InstructGPT). NeurIPS 2022.
Self-supervised
LeCun, Y. (2022). A path towards autonomous machine intelligence. OpenReview.
Course material
BUSN9049 Module 2 — AI Technologies: ML, Deep Learning, NLP, LLMs and Generative AI. Flinders University, 2026.