What is a model in AI?
In one line
Ages 8 to 10: A model is a machine's learned rule: the thing that takes in something new and gives back a guess.
Ages 11 to 13: A model is a function with adjustable numbers, called parameters, and training sets those numbers so its outputs match the examples.
Ages 14 to 18: A model combines a structure chosen by people with parameters learned from data, and it is a simplified stand-in for the world that is trustworthy only as far as it has been tested.
In AI, a model is what a computer builds from examples to turn an input into an output, such as a photo into the word "cat". It is a set of steps plus learned numbers called parameters. Training sets the numbers; using the model on something new is called making a prediction.
Last reviewed: 20 October 2026 6 min read
The idea in one picture
Suppose you want to predict how long a walk will take. You time three walks:
| Distance | Time |
|---|---|
| 1 km | 12 minutes |
| 2 km | 25 minutes |
| 3 km | 35 minutes |
You choose a simple shape for the rule: minutes = something × kilometers. Then you find a "something" that fits well. Across the three walks you took 72 minutes for 6 km, and 72 ÷ 6 = 12, so your rule is minutes = 12 × km.
That rule is a model. Its shape, "a number times the distance", was chosen by you. The 12 was learned from the examples: it is the model's one parameter. Ask about a 4 km walk and the model predicts 12 × 4 = 48 minutes. It cannot know about a steep hill on the way, because no example had one.
How it works, step by step
- Choose the structure. People pick the model's shape: a line, a decision tree or a neural network with a set number of layers.
- Start the parameters at rough or random values.
- Train. Compare the model's outputs with the examples and adjust the parameters to reduce the error.
- Save it. The trained model is its structure plus its learned numbers.
- Predict. Give it a new input and it produces an output.
- Test it on new examples and count its mistakes.
A little deeper (ages 11 to 13)
People mix up model and algorithm. The training algorithm is the procedure that adjusts the numbers; the model is what you have afterwards. A cake recipe is not the cake.
Models range from tiny to huge. The walk model has one parameter. A simple classifier that sorts 32 × 32 color pictures into 10 categories has 10 × 3,072 weights plus 10 biases: 30,730 parameters. Large modern models have far more. More parameters let a model capture more detail, but also need more examples to avoid overfitting.
Going further (ages 14 to 18)
In science, a model is a simplified representation of something real. An AI model is the same: it captures patterns in its training data and leaves everything else out. Its predictions are only as reliable as the tests it has passed, on data like the data it will meet.
A model is also not a product. An app or service wraps a model with other parts: rules, filters, a user interface, sometimes other models. When an app gets something wrong, the cause can be in the model, its training data or the parts around it.
Good practice is to publish a short document with a trained model, sometimes called a model card, that states its intended use and reports how it performs across different groups.
Common mix-ups
- Mix-up: "The model is the app."Fact: The app is built around a model, often with other parts.
- Mix-up: "The model knows things like a person."Fact: It stores learned numbers that turn inputs into outputs.
- Mix-up: "The model and the algorithm are the same."Fact: The algorithm trains; the model is the result.
- Mix-up: "A model in AI is like a 3D model or a fashion model."Fact: Same word, different meaning: here it means a learned rule.
Try it (no device, 5 minutes)
Build a step model. Walk 5 steps across a room and count the seconds, then 10 steps. Work out seconds per step: that is your parameter. Predict how long 20 steps will take, then walk it and check. Now ask someone else to walk 20 steps. Does your model work for them? What would you change?
Quick check
Using minutes = 12 × km, how long does the model predict for 5 km?
Answer: 12 × 5 = 60 minutes.
12 × 5 = 60 minutes.In the walk model, which part was chosen by a person and which was learned?
Answer: The shape "a number times the distance" was chosen; the number 12 was learned from the walks.
The shape "a number times the distance" was chosen; the number 12 was learned from the walks.Name two things a model card should report.
Answer: The model's intended use and how well it performs across different groups.
The model's intended use and how well it performs across different groups.Questions
A model is what a computer builds from examples so it can turn an input into an output, such as a photo into a label. It is a set of steps plus learned numbers, and it is used to make predictions about new cases.
People choose the model's structure and collect training examples. Training then adjusts the model's numbers, its parameters, until its outputs match the examples closely, and the model is tested on examples it has never seen.
An algorithm is a step-by-step procedure, such as the one that trains a model. The model is the result of training: the structure plus the learned numbers that turn inputs into outputs.
No. An app or service is built around one or more models, with other parts such as rules, filters and a user interface. A mistake in the app can come from the model or from those other parts.
A model only captures patterns in its training examples. When a new case is unlike anything it learned from, or the examples had gaps or wrong labels, its prediction can be wrong.
Mechna Academy is a city of machines for curious children aged 8 to 18, opening first to ages 8 to 10. In each mission, a child guesses how a machine decides, then opens its Peek Hatch to check, changes one thing and tests it on a case it has never seen. Children learn how AI works this way, and they never chat with an AI.
For grown-ups: join the waitlistPlaces are limited. The waitlist is for parents and guardians 18 and older.