An AI model is a program that has learned patterns from data and uses those patterns to turn new input into a useful output, such as a prediction, a category label, a recommendation or generated text. The model is the trained core of an AI product. It is what actually does the “thinking” part of the work.
Key takeaways
- A model is created by running a training algorithm on data.
- Once trained, a model is stored as a file of learned numbers called parameters or weights.
- Using a trained model on new data is called inference.
- A model on its own is not a product. It becomes useful inside an AI system or AI application.
An AI model explained in simple terms
Imagine a recipe that writes itself. You give it thousands of cakes and tell it which ones tasted good. Over time it works out the right mix of flour, sugar and heat. The finished recipe is the model. The process of working it out is training. Baking a new cake with the recipe is inference.
In technical terms, a model is a mathematical function with many adjustable settings. Training nudges those settings until the function’s outputs match the examples. A small model might have a few hundred settings. Large language models have billions.
How is an AI model created?
- Define the task. For example: “Is this email spam?” or “What is the next word?”
- Collect and prepare data. Clean it, remove duplicates and, for many tasks, add labels.
- Choose an architecture. This is the model’s shape, such as a decision tree or a neural network.
- Train. An algorithm repeatedly compares predictions with correct answers and adjusts the parameters to reduce error.
- Evaluate. Test on data the model has never seen, to check it generalises rather than memorises.
- Deploy and monitor. Put it into use, watch performance and retrain when results drift.
Main types of AI models
| Type | What it does | Example |
|---|---|---|
| Classification model | Sorts input into categories | Spam vs not spam (AI classification) |
| Regression / forecasting model | Predicts a number | Next month’s sales (AI prediction) |
| Clustering model | Groups similar items without labels | Customer segments |
| Recognition model | Identifies objects, faces, speech or text | Voice-to-text |
| Generative model | Creates new content | Text, images, code |
| Recommendation model | Ranks items for a user | Product suggestions |
Foundation models and large language models
A foundation model is a large model trained on broad data that can be adapted to many tasks. Large language models (LLMs) are foundation models trained mainly on text. Most modern LLMs are built on the transformer architecture, introduced in the 2017 paper “Attention Is All You Need”. Because they are general-purpose, the same model can power chat, summarisation, coding help and search.
AI model vs algorithm vs system
These three words get mixed up constantly. The algorithm is the method used to learn. The model is what the algorithm produces after training. The system is the full setup around the model: data pipelines, interface, rules, monitoring and people. A simple way to remember it: the algorithm is the teacher, the model is the trained student, and the system is the whole school.
How to judge an AI model
- Accuracy: how often it is right on fresh test data.
- Robustness: how well it copes with messy or unusual input.
- Fairness: whether error rates differ across groups of people.
- Speed and cost: how fast and cheap each prediction is.
- Transparency: whether documentation, such as a model card, explains its data, limits and intended use.
Limitations
A model only knows what its data taught it. If the data is old, narrow or biased, the model will be too. Models can also “overfit”, performing well in testing but poorly in real use. And a model’s confidence is not proof it is correct. This is why models need ongoing monitoring, not a one-time launch.