Machine learning is software that finds patterns in data and uses them to make predictions, rather than following rules a programmer wrote. It powers spam filters, fraud detection, recommendations, medical imaging analysis, translation and the language models behind modern AI assistants.
What is Machine Learning?
Traditional software follows explicit rules. A programmer writes: if the transaction is over R10,000 and from a new device, flag it. Every rule has to be thought of and written down.
Machine learning works the other way round. You give the system thousands of examples — transactions labelled fraudulent or legitimate — and it works out for itself which patterns distinguish them. Nobody writes the rules; the system derives them from the data.
That is the whole idea. Everything else is detail about how.
Three broad kinds: supervised learning, trained on labelled examples — most practical applications are this; unsupervised learning, which finds structure in unlabelled data, such as customer segments nobody defined; and reinforcement learning, which learns by trial and reward.
Deep learning is a subset using neural networks with many layers. It is what made image recognition, speech recognition and language models work well enough to be useful, and it is behind essentially every AI advance of the last decade.
Machine learning is a subset of AI, not a synonym. Artificial intelligence is the broad goal; machine learning is the technique that currently dominates it.
Two honest limits, since both are widely oversold:
A model is only as good as its data. Trained on biased data, it produces biased predictions — reliably and at scale. There are documented cases of hiring and lending systems disadvantaging groups because the historical data they learned from did.
It predicts; it does not understand. A model that classifies X-rays has no concept of a lung. It found statistical patterns, and it can fail in ways that make no sense to a person — which is why high-stakes uses keep a human in the loop.
What is Machine Learning used for?
Things you already use daily:
- Spam filtering in email.
- Recommendations — Netflix, Spotify, Takealot, YouTube.
- Search ranking.
- Predictive text and autocorrect.
- Face recognition to unlock a phone or tag photos.
- Voice assistants and dictation.
- Navigation and traffic prediction.
- Language models behind AI chat assistants.
In business:
- Fraud detection in banking and card payments — one of the earliest and most successful uses, and heavily used by South African banks.
- Credit scoring, demand forecasting and dynamic pricing.
- Customer churn prediction and chatbots.
- Document processing — reading invoices and forms.
- Predictive maintenance, flagging equipment likely to fail.
In medicine and science:
- Medical imaging — detecting tumours, diabetic retinopathy and TB on chest X-rays. TB screening by machine learning is genuinely significant in South Africa, where it extends screening capacity where radiologists are scarce.
- Drug discovery and protein structure prediction.
- Genomics.
Elsewhere: agriculture (crop and disease detection from drone imagery, used in South African farming), wildlife conservation (identifying poachers and counting animals from camera traps), weather forecasting, and speech-to-text.
Common tasks you can do with Machine Learning
Classification. Sort things into categories — spam or not, fraudulent or legitimate. The most common practical task.
Regression. Predict a number — house price, next month’s sales.
Clustering. Group similar items without predefined categories.
Recommendation. Suggest items based on behaviour and similarity to others.
Natural language processing. Sentiment analysis, summarising, translation, extracting information from documents.
Computer vision. Identify objects, read text from images, detect production defects.
Forecasting and anomaly detection. Predict a time series, or flag what does not fit the normal pattern.
Practically, you rarely build from scratch. Most work uses pre-trained models and adapts them, or calls a cloud API from AWS, Google Cloud or Azure. For many business problems, an API call is the whole solution — and building a custom model is the wrong instinct.
Who Machine Learning is useful for
Data scientists and ML engineers, obviously — a well-paid and in-demand specialisation.
Software developers adding intelligent features without becoming specialists, usually through APIs.
Analysts moving from describing what happened to predicting what will.
Business owners, mostly as consumers of ready-made tools rather than builders. For most small businesses, the right move is buying software that already uses machine learning, not hiring someone to build a model.
Researchers, across nearly every field.
Students. Python is the standard language, and demand for these skills in South Africa is genuine and growing, including in remote roles paying internationally.
Anyone worried about being replaced. Worth being straightforward: machine learning automates tasks, not usually whole jobs, and it is best at repetitive pattern-matching on large volumes of data. Judgement, physical work, relationships and situations with little historical data remain hard for it.
When to choose Machine Learning instead of a related tool
| Situation | Better choice |
|---|---|
| Rules are clear and stable | Ordinary programming — simpler, cheaper, explainable |
| Patterns are complex, data plentiful | Machine learning |
| Understanding what already happened | Analytics and dashboards |
| Repetitive clicking between systems | Workflow automation |
| Text generation, summarising, drafting | A large language model |
| Very little data | Traditional statistics or rules |
| A decision needing an explanation | Rules or simple models — complex ones are hard to explain |
Choose machine learning when you have substantial relevant data, the pattern is too complex to write down, some error rate is acceptable, and the problem repeats often enough to justify the work.
Choose something else when the rules are simple enough to write, you lack data, errors are unacceptable, or the decision must be explainable — which matters legally in lending and hiring.
See the AI and automation hub, and the guide to Python.
Related software and productivity guides
Frequently asked questions
What is machine learning used for? Finding patterns in data to make predictions — spam filtering, fraud detection, recommendations, medical imaging, translation, forecasting and the language models behind AI assistants.
How is it different from ordinary programming? Ordinary software follows rules a programmer wrote. Machine learning derives the rules itself from examples.
Is it the same as AI? No. AI is the broad goal; machine learning is the technique that currently dominates it. Deep learning is a subset of machine learning.
Do I need to be a programmer? To build models, yes — usually Python. To use machine learning, no; most business software already includes it.
How much data do I need? More than most people expect. Simple problems may work with thousands of examples; image and language tasks typically need far more, which is why pre-trained models are usually adapted rather than built from scratch.
Can it be biased? Yes, and reliably so. A model trained on biased data reproduces that bias at scale. This is documented in hiring and lending systems.
Will it take my job? It automates tasks more than whole jobs, and it is best at repetitive pattern-matching over large data volumes. Judgement, physical work and relationship-based work remain hard for it.
What should I learn first? Python, then the basics of statistics, then a library like scikit-learn before moving to deep learning frameworks.