AI glossary
AI terms explained without the fog.
AI writing often becomes hard to read because simple ideas are hidden behind technical terms. This glossary gives short, useful definitions and links readers back to deeper guides when a concept needs more context.
How to use this glossary
Read a term, then ask what role it plays in a real AI workflow. Some terms describe the model, some describe the data, some describe risk, and some describe the way people use the tool. That distinction helps readers avoid treating every AI word as the same kind of claim.
When an AI product uses one of these terms, look for the practical meaning. If a page says "AI-powered," ask what input the system uses and what output it produces. If it says "accurate," ask how accuracy was measured. If it says "safe," ask what risks were tested and who reviews mistakes.
Artificial intelligence
Software designed to perform tasks that normally need human-like pattern recognition, language understanding, prediction, or decision support.
Machine learning
A way of building systems that learn patterns from examples instead of relying only on hand-written rules.
Model
The trained system that receives an input and produces an output, such as a prediction, ranking, summary, or generated draft.
Generative AI
AI that creates new text, images, code, audio, or other content from instructions and learned patterns.
Prompt
The instruction, question, context, or example a user gives an AI tool to guide its response.
Training data
The examples used to teach a model patterns before it is used on new inputs.
Label
A tag or answer attached to training examples, such as spam/not spam or safe/unsafe.
Validation data
Data used during development to tune a model and compare versions.
Test data
Separate examples used to estimate how well a model handles data it has not trained on.
Bias
A pattern that can lead to unfair, incomplete, or unreliable outcomes, often because data or design choices are uneven.
Hallucination
A confident AI answer that includes false, unsupported, or invented information.
RAG
Retrieval-augmented generation: a method that lets an AI system retrieve outside information before generating an answer.
Embedding
A numeric representation of text, images, or other data that helps software compare meaning or similarity.
Fine-tuning
Additional training that adapts a model toward a narrower task, style, or domain.
Guardrail
A rule, filter, workflow, or review step used to reduce unsafe or unwanted AI behavior.
Human in the loop
A workflow where people review, approve, or correct AI output before it is used.
Token
A chunk of text a language model processes. Tokens may be words, parts of words, or punctuation.
Context window
The amount of text or information an AI model can consider at one time.
Inference
The moment a trained model is used to produce an answer, prediction, classification, or generated output.
Classification
A task where a system sorts input into categories, such as spam or not spam.
Regression
A task where a system predicts a number, such as demand, price, time, or probability.
Recommendation system
Software that ranks or suggests items based on patterns in user behavior, item details, or similarity.
Confidence score
A number that may describe how strongly a system prefers an answer, though it should not be treated as proof of truth.
Grounding
Connecting an AI answer to specific documents, data, or sources so the response is easier to verify.
Evaluation
The process of testing whether an AI system performs well enough for the task and risk level.
Benchmark
A standard test used to compare systems, though real-world performance may differ from benchmark results.
Overfitting
When a model performs well on familiar examples but struggles with new or different cases.
Dataset shift
When real-world data changes after training, causing a model's performance to weaken.
Explainability
How clearly people can understand why a model produced a result or which factors influenced it.
Automation
Using software to complete a task with little or no human action after setup.
Augmentation
Using AI to support a person while the person keeps review, context, and final responsibility.
Synthetic data
Artificially generated data used for testing, training, or privacy protection, but still requiring careful validation.
Personally identifiable information
Information that can identify a person, such as names, contact details, IDs, or combinations of details.
Model card
A document that describes a model's purpose, limits, data, evaluation, and intended use.
Data sheet
A document that describes where a dataset came from, what it includes, and what limits or risks it has.