AI Glossary: 30 Common Terms in Plain English

New vocabulary is one of the biggest barriers to using AI well. A handful of terms explain most of what you will hear in a meeting, a product demo, or a vendor pitch. This glossary collects thirty of them, grouped by what they actually describe, so the definitions support each other instead of standing alone.

The entries here are deliberately plain. They favor usefulness over precision, and each one notes where the everyday explanation leaves something out. For the fuller picture, see What Is Generative AI?, How AI Assistants Work, and AI Models vs. AI Applications.

How to use this glossary

You will get the most out of these definitions if you treat them as connected rather than isolated. The following short routine works well for anyone new to the subject.

1. Start with the core concepts. Learn model, application, prompt, and token first; most other terms depend on them.
2. Move to working with models. Understand training, inference, and context window to see why results vary.
3. Read the risks section slowly. Hallucination, bias, and prompt injection explain most real-world failures.
4. Finish with evaluation. Terms like benchmark and ground truth help you judge claims without trusting them.
5. Keep the caveat in mind. Every entry is a simplification; verify the specific behavior of the tool you actually use.

Core concepts

These are the words that show up in almost every AI conversation. Getting them straight removes most of the confusion early.

Artificial intelligence (AI)

A broad label for software that performs tasks we associate with human thinking, such as recognizing patterns, understanding language, or making recommendations. It covers many techniques, and not all of them generate text.

Generative AI

AI that produces new content rather than only classifying or scoring existing content. Output can be text, images, audio, or code. The "generative" part is what makes it flexible and what allows it to sound convincing even when wrong.

Model

The trained engine that turns input into likely output. A model has no interface and no memory of you. Its capability, knowledge cutoff, and cost per use are all properties of the model itself.

Large language model (LLM)

A type of model trained on very large amounts of text to predict the next piece and, from that, to write and reason about language. Chat assistants are usually built on an LLM, but the model is only one part of the product.

Application

The product you actually open and use. It wraps the model with an interface, memory, search, file handling, and policies. Most features people praise or complain about belong to the application, not the model.

Prompt

The instruction you give the model. A prompt shapes output the way a brief shapes a contractor's work: more context, a target audience, and a requested format reliably improve the result.

Token

A small chunk of text — often a word or part of a word — that the model processes as a unit. Tokens explain length limits, cost, and why models can stumble on exact character-level details such as counting letters.

Parameters

The numbers inside a model that are adjusted during training and that store what it has learned. When people say a model is "larger" or "smaller," they usually mean it has more or fewer parameters, though size alone does not determine quality.

Working with models

These terms describe the two phases of a model's life — being built and being used — and the settings that affect what you get back.

Training

The process in which a model learns from examples and adjusts its parameters. A vendor usually handles base-model pretraining before releasing a tool. Further training, such as fine-tuning, can adapt a model for an organization; ordinary chat use does not itself update the model's parameters.

Inference

Everything that happens when you actually use a trained model: your input goes in, a prediction comes out. Inference is fast and, by default, forgetful — the model does not learn from your conversation.

Fine-tuning

Further training of an existing model on a smaller, specific set of examples to nudge its behavior or style. It is often confused with retrieval, which supplies information at use time instead of changing the model.

Context window

The total amount of text a model can consider at once, including your instructions, prior messages, and its own reply. Think of a desk that only holds so many pages; when it fills, earlier material is dropped or summarized.

Temperature

A setting that controls how random the model's choices are. Low values give more predictable, focused output; high values give more varied, creative output — and more drift from the point.

Embedding

A way of representing text as a list of numbers that captures meaning, so that similar ideas sit close together. Embeddings power features like semantic search and document retrieval.

Retrieval-augmented generation (RAG)

A pattern where an application fetches relevant documents and passes them to the model along with your question. It reduces reliance on the model's memory, but it does not guarantee that the retrieved source is accurate or correctly read.

System prompt

Instructions set by the application, not by you, that come before your message and shape the assistant's role and rules. This is one reason two products on the same model can behave differently.

Risks and limits

These terms describe the ways AI systems fail. Understanding them is the most practical skill in the field.

Hallucination

Output that is fluent, confident, and false — invented facts, citations, or numbers delivered in the same tone as true ones. It reflects how the model predicts plausible text rather than checking facts, and it is a structural property, not a glitch.

Bias

Systematic skew inherited from training material, which can lead to unfair or unbalanced output about groups of people. It matters most in decisions that affect people, where unreviewed output is a genuine risk.

Knowledge cutoff

The point after which a model has no built-in knowledge, because it stopped training. Without a connected search or retrieval step, the model may answer anyway rather than admitting the gap.

Prompt injection

An attack where hidden instructions in text the model reads — a web page or a document — hijack its behavior and make it follow the attacker's intent instead of yours. It is a reason to treat any untrusted input with caution.

Data leakage

The unintended exposure of sensitive information through a model or application. It can happen when confidential text is stored or surfaced in ways you did not expect, which is why data terms deserve attention before you paste anything.

Alignment

The effort to make a system behave in line with human intentions and values. In practice it is imperfect: alignment techniques reduce some harmful behavior without eliminating errors or bias.

Guardrails

Application-level rules that restrict or filter what a model can do, such as blocking certain topics or actions. Guardrails explain many refusals that look like model limits but are actually policy choices.

Evaluation and data

These terms help you judge claims about AI quality instead of taking them at face value.

Benchmark

A standardized test used to compare models on tasks like reading comprehension or reasoning. Useful for rough comparison but easy to overstate, because high scores do not guarantee good performance on your specific work.

Ground truth

The verified, correct answer used as a reference when checking whether a model or system is accurate. Without ground truth, "looks right" is the only standard, and it is a weak one.

Red teaming

Deliberately probing a system for weaknesses — harmful outputs, unsafe behavior, or ways to bypass restrictions — before others find them. It is a form of structured skepticism applied to AI.

Human in the loop

A design in which a person reviews or approves AI output before it takes effect. For anything that affects money, safety, or people's standing, keeping a human in the loop is a baseline safeguard, not a nice-to-have.

Latency

The delay between making a request and receiving a response. Longer responses and larger models often mean more latency, which shapes whether a tool feels practical in a live conversation.

Multimodal

Describes a model that handles more than one kind of input or output, such as text and images together. It broadens what you can ask for while adding new places for errors to hide.

Open-weight model

A model whose trained parameters are published for others to download and run. It offers more control and can be hosted privately, but "open" says nothing about accuracy or safety, which still depend on the model and how it is used.

Limitations of this glossary

Plain-language definitions trade precision for clarity, and that trade has real limits. A term like hallucination is easy to grasp but covers a range of causes that researchers distinguish carefully. Temperature is described here as a randomness dial, while in practice it interacts with other sampling settings you may never see.

Definitions also drift as products evolve. A term that meant one thing two years ago can quietly change meaning as vendors repackage features. And a definition tells you how something generally behaves, not how a specific tool behaves today — which is why current documentation matters whenever a detail is important.

Above all, no vocabulary list replaces verification. Knowing what a word means does not tell you whether a particular output is correct, and the most confident-sounding terms are often the ones that mask the most uncertainty. For the practical side, see the fact-checking checklist and our guide to limitations and verification.

Conclusion

Thirty terms cover most of what you will encounter in everyday AI use, and they cluster naturally: core concepts, working with models, risks and limits, and evaluation. Learning them in that order builds understanding instead of scattered trivia, and it lets you follow a conversation without overstating what the technology actually does.

The vocabulary is a foundation, not an ending. Put it to work through AI Basics, take care with sensitive material in Protecting Confidential Information, and treat fairness as a design requirement rather than an afterthought with our responsible AI guide.