Quest 2 of 15
How AI Learns from Data
Understand the core AI loop: data becomes patterns, and patterns support predictions or classifications. You will learn why training and testing use different examples, without needing mathematics. This matters because a model trained only on one kind of email may fail when it sees another language or writing style. You will also recognise how incomplete or biased data can create harmful outcomes.
Start here
When people say a computer “learns,” they do not mean it studies, understands, or becomes wise like a person. They mean a training process changes the model’s numerical settings so it becomes better at producing the expected answers on examples. This lesson follows that process one step at a time.
Big idea
A machine-learning model learns statistical relationships from examples. The examples must represent the real task, labels must be dependable, and evaluation must use data the model did not train on. A high score is useful evidence, but it is never a guarantee that the model will work safely for every person, place, or future situation.
How machine learning is built
Click each stage — no coding required to understand the flow.
Gather many examples: emails, photos, sensor readings — with labels where needed.
Learn one idea at a time
Read, explore, then mark each idea when you can explain it.
Idea 1 of 10
Data simply means recorded information. One row might describe an email, one photograph, one card payment, or one hour of readings from a pump. Useful data includes the details related to the task. For spam detection, that might include message text, sender history, links, and whether a person previously marked the message as spam.
Choose a deep dive
Open the topics you want to explore. The detail stays folded until you need it.
Deep dive 1Worked example: teaching a system to recognise crop disease
A team collects leaf photographs from different farms, seasons, phone cameras, and lighting conditions. Plant specialists label each image after inspection. The team trains on most images but keeps a separate test set. It then checks not only overall accuracy but also errors on early-stage disease, dusty leaves, and lower-quality phones. Before farmers act, uncertain cases are referred to an extension worker. The model assists screening; it does not replace field diagnosis.
Deep dive 2Accuracy can hide important mistakes
Suppose only 1 in 100 transactions is fraudulent. A useless model that predicts “not fraud” every time is 99% accurate, yet it catches no fraud. Teams must choose measures that match the consequence: how many real fraud cases were caught, how many legitimate customers were disturbed, and what losses followed. The same lesson applies to health screening, safety alarms, and quality inspection.
Deep dive 3Data is about people and processes, not just files
Every dataset has a history. Somebody decided what to record, which cases were missing, how labels were assigned, and who could access the result. Understanding that history—often called data provenance—helps reveal limits. A clean spreadsheet can still be unsuitable if it excludes rural users, contains outdated policy decisions, or was collected for a different purpose.
Deep dive 4A small training set can teach the wrong lesson
Suppose a crop-photo model is trained only on maize leaves photographed in bright sunshine. It may work well in that narrow set but fail on cloudy-day photos or leaves from another district. The missing situations are not a minor detail: they are where people will discover whether the tool is dependable.
Deep dive 5Accuracy can hide harm
A model that correctly labels 90 out of 100 messages may sound excellent. But if all ten mistakes are urgent health messages marked as ordinary, the result is dangerous even with 90% accuracy. Builders must ask which mistakes happen and who carries the cost, not only read one big percentage.
Deep dive 6Case study: teaching a model to read handwriting
Imagine a project to digitise handwritten clinic records. Volunteers label ten thousand scanned words with their correct text. The model learns quickly on neat handwriting but struggles with rushed notes, faded ink, and abbreviations like 'BP' for blood pressure. The team responds not by changing the algorithm but by adding more examples of exactly the difficult cases: rushed notes, poor scans, and common abbreviations, each carefully labelled. Performance improves because the training data now resembles real life. The lesson: when a model fails, the most common fix is better, more representative examples — not a cleverer formula.
Deep dive 7Where labels come from, and why it matters
Labels feel objective, but a person or process chose each one. Was a loan application 'bad' because the borrower truly could not pay, or because the bank's collections process failed them? Was a photo labelled by an expert, a volunteer, or another algorithm? When two people disagree about a label — is this message rude or just direct? — the model inherits that ambiguity. Good teams write labelling guidelines, measure how often labellers agree, and audit samples. When you evaluate any AI product, asking 'who labelled the training data and how carefully?' tells you a great deal about how much to trust it.
Deep dive 8Examples teach the model
Machine learning needs many labelled examples. If the examples are incomplete or unfair, the model can copy those problems into new decisions.
One-minute challenge
Connect this lesson to real life
Name one situation where this idea could help, and one thing a person should still check.
Explore a real-world example
Use the arrows to connect the idea to a visible situation.
Photo example
Example: pattern-finding at scale
Behind a simple chat answer may be large computer systems trained on huge text collections. Quality and diversity of that data shape what the model can do well.
Key terms
Tap a term to flip and read the definition.
Optional further learningFree textbooks and trusted online resources
These sources informed the course structure. Use them to revisit a concept or study it in more depth.
Ready check
Tick each idea only when you could explain it without looking back.
Ready for practice? Walk through the ML pipeline diagram, then label the practice emails.
Extra context (audience, logistics, curriculum notes)
Built for: Continues from Part 1; still zero coding.
Formats: Pipeline diagram · Matching activity · Quiz
Introduction to AI — Part 2: data, labels, training vs testing.
Next up
Ready for the next part?
When you've finished the reading, inline exercises, and knowledge check for this part, check the box to continue.