AI Concepts A–Z

This is a plain-English concept index for HowAIModelsWork.com. Use it like a practical glossary: each concept gives you a short explanation and a link to a deeper article or guide.

A

AI agent

An AI system that may plan steps, use tools and act across a task.

Read more →
AI alignment

The attempt to shape AI behavior toward intended human goals and rules.

Read more →
AI model

A trained system that produces outputs from learned patterns.

Read more →
Attention

A mechanism that helps a model weigh which parts of the input matter most.

Read more →
AI audit

A structured review of an AI system’s data, measurements and outcomes to find uneven behaviour, weak tests or hidden risks.

Read more →

B

Bias

Systematic patterns in AI outputs that can come from data, labels, objectives, evaluation choices or the way a system is used.

Read more →
Batch

A smaller group of training examples processed together during one training step.

Read more →
Benchmark

A standard set of tasks or measurements used to compare model performance, with limits on what the score proves.

Read more →

C

Chain-of-thought prompting

A prompting style that asks the model to work through steps.

Read more →
Chunking

Splitting documents into pieces so retrieval systems can search them more easily.

Read more →
Computer vision

AI systems that process images and visual information.

Read more →
Context window

The amount of text and information the model can consider at one time.

Read more →
Checkpoint

A saved version of a model during training that engineers can test, compare or resume from.

Open guide →

E

Embeddings

Number-like representations that help AI compare meaning.

Read more →
Evaluation

The process of testing a model on chosen tasks, groups and failure cases before or after release.

Open guide →

F

Fine-tuning

A later training step used to adjust a model for a task, behavior or style.

Read more →
Function calling

A way for AI to connect a response to a tool or external action.

Read more →
Fairness

A system-level question about how data, objectives, thresholds and consequences affect different people or groups.

Open guide →

G

Generative AI

AI that creates text, images, code, audio or other outputs.

Read more →
Grounding

Connecting an AI answer to source material or external information.

Read more →

H

Hallucination

A fluent AI answer that may be wrong, unsupported or invented.

Read more →
Heatmap

A visual overlay showing which image regions influenced an AI result, without revealing a human-like thought process.

Read more →

L

Latent space

An internal numerical space where a model represents patterns and relationships between visual or other features.

Read more →
Loss

A number that measures how far a model’s prediction is from its training target and guides parameter updates.

Read more →

M

Mixture of Experts

A model design that activates selected expert parts instead of using everything equally.

Read more →
Model collapse

A risk that can appear when models learn too much from AI-generated material.

Read more →

P

Proxy variable

A feature that can carry information about another characteristic even when that characteristic is not collected directly.

Open guide →

R

RAG

Retrieval-augmented generation: when AI looks things up before answering.

Read more →
Reasoning model

A model designed to spend more effort on multi-step problems.

Open guide →
Retrieval

Finding source material before generating an answer.

Read more →
RLHF

Reinforcement learning from human feedback, used to shape model behavior.

Read more →

S

Sampling

How a model chooses one possible next token from many possible options.

Read more →
Synthetic data

Data generated artificially, sometimes by AI systems themselves.

Read more →
Surrogate model

A faster learned approximation that imitates part of a slower scientific simulation.

Read more →

T

Temperature

A setting that affects how predictable or varied an AI answer may be.

Read more →
Token

A small piece of text that an AI model reads or predicts.

Read more →
Transformer

A model architecture behind many modern language models.

Read more →

V

Vector database

A database designed to search by meaning instead of exact keywords.

Read more →
Voice AI

AI systems that process or generate spoken language.

Read more →
Visual tokens

Small numerical image regions or feature groups that let a visual model process a picture in manageable pieces.

Read more →
Validation

The work of checking whether a promising model result survives independent tests, experiments or real-world use.

Open guide →

Want a guided route?

If you prefer a reading order instead of an index, use the AI Learning Paths page.

Open learning paths →

Readers Also Read

Why AI Sometimes Refuses Harmless Questions

Why Voice AI Mishears Certain Words

Why AI Voice Assistants Pause, Hesitate or Interrupt

Why AI Sometimes Chooses Caution Over Precision