How AI learns
Show a child enough dogs, and one day they point at a dog they've never seen. AI learns the same way, from examples. That's training. Then it predicts.
First, what do you think?
Where do the patterns of AI come from? Have a guess.
How does AI learn?
The whole story in three sentences. AI learns from lots of examples. From those examples, it finds patterns. After that, it uses the patterns to predict for something new.
Think of a toddler learning what a dog is. You point and say "dog". A big dog, a small dog, a dog in a book. You point at a cat and say "no, that's a cat". One day the toddler sees a dog they've never seen. "Dog!" Nobody gave them a rule. They found the pattern.
AI learns in a similar way. Three words help.
- Data are the examples. Photos, emails, sentences, prices. Often with the right answer attached: dog or not dog, spam or not spam.
- Training is the learning. The computer goes through the examples and finds the patterns.
- The model is the result: the learned patterns, packed into software.
Training and using are two separate moments.
- Training happens first, often once, and it's a lot of work. Big models learn from a huge amount of data.
- Using comes after. When you type a question or upload a photo, the model doesn't learn. It predicts with what it already learned.
A first small step: an app recognises something? Ask where its examples came from.
What the predictor does
The predictor can only be as good as its examples. That's the rule.
An example. Imagine you train a model to recognise apples. You only show it red apples. Then you show it a green apple. It may predict "not an apple". Nothing went wrong in the model. Its examples never had a green apple.
That explains a lot.
- More examples help. With more examples, the patterns get more reliable.
- Varied examples help more. Red, green and yellow apples. Big and small. In a bowl and on a tree.
- Wrong examples teach wrong patterns. If the examples contain mistakes or unfair patterns, the model learns those too. Models share the biases in their data.
The predictor can only be as good as its examples.
Try it yourself
Be the trainer. Choose the examples, and see what the predictor does with them.
Everything runs in your browser. Nothing you type is sent anywhere.
In practice
Three people meet a predictor that learned from examples. The examples explain the result.
Case 1The face that wasn't found
Lotte (16) uses a photo app that groups pictures by face. It finds her friends, but keeps missing her grandmother.
Think the app is broken.
The app works as trained. Maybe it saw few faces like her grandmother's. Maybe there are only two photos of her. The predictor has little to go on.
If the app lets you, add her grandmother's name to a few photos yourself.
You give the predictor more examples of exactly the face it missed.
Case 2The predictor that learned the wrong thing
Ahmed (41) works at a company that tested a tool to sort job applications. The tool learned from the people hired in the past ten years.
Trust it: a computer can't be unfair.
The tool learned the patterns of the past. Did the company use to hire mostly one kind of person? Then the tool predicts that kind is best.
Ask what data a tool learned from before you use it. Keep a person checking the result.
Unfair examples give unfair predictions. A person can spot what the pattern hides.
Case 3"Does it learn from me now?"
Marja (59) asks a chatbot for a recipe. Then she wonders: does it now remember her taste forever?
Think every chat makes the model learn.
Training and using are separate. While you chat, the model predicts with what it already learned. It doesn't train on your words while you type.
Know that a company may use chats later to train a new version. Check the settings of the chatbot you use.
You know the difference between using and training. So you know where your words can end up.
By situation
You learn by practising with examples, too. But you can ask why. The predictor can't. It only knows which patterns showed up often. That's why it can be confidently wrong about things that were rare in its examples.
Before your team uses an AI tool, ask: what did it learn from? Does that look like our customers, our language, our work? A tool trained on other situations predicts for those situations.
Apps learn from what millions of people do. That's why suggestions fit most people, but not always you. Your choices are examples, too. What you click on helps shape what the next version learns.
Myths and mistakes
| Claim | True? | Why |
|---|---|---|
| A computer is always fair | No | A model learns the patterns in its data, unfair ones too. |
| AI learns from every chat while you type | No | Training and using are separate. Some companies do use chats later to train a new version. |
| More data always fixes everything | No | More of the same examples doesn't help with what's missing. Variety matters. |
| The model contains all its examples | No | It contains patterns learned from them, not a copy of every example. |
Related topics
Read first
- What is AI?rules versus learning from examples.
Next
- Kinds of AIwhat different predictors predict.
Related
- Training and the knowledge cutoffwhy a chatbot doesn't know yesterday's news.
- Privacy and your datawhether your chats are used for training.
Sources
- explainerGoogle for Developers. What is Machine Learning?.Machine learning trains software, called a model, to make useful predictions from data instead of hand-written rules. Four kinds: supervised, unsupervised, reinforcement learning and generative AI. Examples: spam, cats in photos, weather.
- explainerGoogle for Developers. LLMs: Fine-tuning, distillation, and prompt engineering.A base model is a platform, not a finished product; fine-tuning on examples for one task improves it. A prompt changes the output, not the model. Models share the biases in their data and cost a lot of computing power.
- documentationAnthropic. Is my data used for model training?.Chats on Free, Pro and Max are used for training if you allow it. You change that with the Model Improvement setting. Incognito chats are never used.
- documentationGoogle. Gemini Apps Privacy Hub.Some chats are read by human reviewers, so don't type anything confidential. With Keep Activity on, chats can train Google's models; turn it off and they're kept 72 hours.
Test
You get 5 questions about the whole page. Get 4 right, and you've mastered the topic.
About the tools in this course. Every tool is described in the same way, in alphabetical order, without a recommendation. Names, prices and features change often: the maker's own site is always right. This site was built with the help of Claude, one of the tools described here.