An AI system receives input, transforms it with an algorithm or trained model, and produces an output such as a prediction, classification, recommendation, or generated response. In machine learning, training adjusts many numerical weights so the model makes fewer errors on examples. Inference uses the trained weights on new input.
How machine learning works in six steps
- Define a task. Decide what the system should predict or produce and how success will be measured.
- Collect and prepare data. Examples may contain text, images, sound, sensor readings, labels, or feedback. Their quality and coverage shape the result.
- Choose an architecture. Engineers select the structure of the model: for example, a decision tree, neural network, or Transformer.
- Train the model. The system makes predictions on examples, measures error with a loss function, and adjusts internal weights to reduce that error.
- Evaluate it. Developers test the model on data it did not train on and examine performance, reliability, bias, security, and failure cases.
- Run inference. A deployed application sends new input through the trained model and turns its output into a useful action or response.
Do not confuse these four parts
Training data
The examples used to adjust the model. Data is not the same thing as the final model.
Architecture
The designed structure that determines how information can flow and be transformed.
Learned weights
The numerical parameters changed during training. They encode statistical patterns, not a database of perfectly retrievable facts.
Application
The software around the model: interface, retrieval, tools, safety checks, storage, and business rules.
What is a neural network?
A neural network is a model made of connected layers of numerical operations. Each layer transforms a representation into another representation. During training, an optimization process adjusts weights across those layers. “Neural” is an historical analogy; artificial neural networks are mathematical systems, not small digital brains.
What is the difference between rules and learning?
In a rule-based system, people explicitly write conditions such as “if this happens, do that.” In machine learning, people define the task, data, architecture, and training process, while the model learns parameters from examples. Many real applications combine both: a learned model makes a prediction and ordinary software enforces permissions, thresholds, or workflows.
Why can AI be confidently wrong?
A model optimizes for a training objective, not for human truth in every situation. Its data can be incomplete or biased; a new input can differ from the training examples; and a generative model selects plausible continuations rather than consulting an internal fact checker. Reliable applications therefore add evaluation, grounding, citations, monitoring, and human review appropriate to the risk.
Common questions
Does an AI understand what it says?
AI can model complex relationships and produce useful behavior, but that does not prove human-like understanding, consciousness, or intention. Treat outputs according to demonstrated performance, not human-sounding language.
Does AI copy its training data?
Training generally adjusts parameters to capture patterns rather than storing a conventional library. However, models can sometimes memorize and reproduce parts of training data, which is one reason privacy, copyright, and testing matter.
Is an algorithm the same as an AI model?
No. An algorithm is a procedure. A trained model is an artifact produced by a learning procedure, with parameters fitted to data.
Primary and institutional sources
Research papers explain mechanisms; institutional resources provide guidance. Manufacturer pages describe their own products and are not independent evaluations.
- ImageNet Classification with Deep Convolutional Neural Networks (2012) — Krizhevsky, Sutskever & Hinton · NeurIPS
- Attention Is All You Need (2017) — Vaswani et al.
- Artificial Intelligence resources — NIST
Continue with the companion in ChatGPT
Explore museum themes, ask questions and test your understanding with quizzes in ChatGPT. The companion and the game keep separate progress.
Discover the ChatGPT companionContent reviewed: 6 September 2026. Navigation and resources updated: 1 October 2026. Project method and limits.