The AI glossary: 55 terms in plain English

The basics

This is the reference version: 55 terms, definition first, detail second. If you want the gentler tour of just the words that matter in your first week, read AI terms explained instead and come back when a term here sends you.

Artificial intelligence

Artificial intelligence is software that performs tasks which normally require human judgment, such as writing, reasoning, or recognizing what is in an image.

In everyday use in 2026, "AI" almost always means the chat assistants built on large language models rather than the broader academic field.

Machine learning

Machine learning is the technique of teaching software by showing it examples rather than programming it with rules.

Every modern AI assistant is built this way: the behavior comes from patterns in training data, which is why the same model can be brilliant on common tasks and shaky on rare ones.

Model

A model is one specific trained AI system, the artifact you actually send a prompt to.

GPT, Claude, and Gemini are model families; a family contains individual models with different sizes, speeds, and prices. Choosing between them per task is the core skill of using AI well.

LLM (large language model)

An LLM, or large language model, is a model trained on enormous amounts of text to predict what text should come next.

That single trick, done at scale, is what produces answers, drafts, code, and summaries. The "large" refers to both the training data and the parameter count.

Chatbot

A chatbot is the conversational interface wrapped around a model, the thing with the message box.

The distinction matters when comparing products: two chatbots can wrap the same model and feel different, and one chatbot can offer many models, which is how Whizi works.

Prompt

A prompt is everything you send the model: the question, the instructions, and any text or files you include.

Answer quality tracks prompt quality closely. Context, constraints, and the output format you want are the three things beginners most often leave out.

Token

A token is the small chunk of text a model actually reads and writes, roughly three quarters of an English word on average.

Pricing, speed, and context windows are all measured in tokens, which is why the word appears on every AI pricing page. A 1,000-word document is roughly 1,300 tokens.

Context window

A context window is the maximum amount of text, measured in tokens, that a model can consider at once.

Your prompt, the conversation so far, attached documents, and the answer all share it. The context window comparison table lists the current figure for every major model, and the explainer covers why it explains most "the AI forgot" complaints.

Parameters

Parameters are the internal numbers a model learned during training, and the usual measure of its size.

Counted in billions, more parameters loosely means more capability and more cost. It is a poor shopping metric on its own; benchmarks and your own tests beat parameter counts.

Response

A response is the model's output for one prompt, also called a completion.

Responses are generated fresh every time, which is why regenerating the same prompt gives different wording and sometimes a different answer.

Talking to a model

System prompt

A system prompt is the standing instruction a model receives before your message, setting its role, tone, and rules.

Every product has one even when you cannot see it. When a chatbot "has a personality", the system prompt is usually where it lives.

Prompt engineering

Prompt engineering is the practice of writing prompts deliberately to get reliably better output.

Despite the grand name, most of it is ordinary clarity: state the goal, give context, show an example, name the format. The beginner guide covers the techniques that actually move results.

Zero-shot

Zero-shot means asking a model to do a task without giving it any examples.

Modern models handle most everyday tasks zero-shot. The term survives mostly as a contrast with few-shot prompting.

Few-shot

Few-shot means including a handful of worked examples in the prompt so the model copies the pattern.

It is the single cheapest reliability upgrade for formatting-sensitive work: two or three examples of input and desired output beat a paragraph of description.

Chain of thought

Chain of thought is a model working through intermediate steps before giving its final answer.

Asking a model to reason step by step measurably improves hard problems, and reasoning models now do a version of it internally by default.

Temperature

Temperature is the setting that controls how much randomness a model adds when choosing each next token.

Low temperature gives consistent, conservative output; high temperature gives variety and occasional nonsense. Chat products usually pick a middle value for you.

Streaming

Streaming is the model sending its answer token by token as it generates, rather than all at once when finished.

It is why answers appear to type themselves, and why a bad answer can be stopped early instead of waited out.

Regenerate

Regenerate means asking for a fresh response to the same prompt.

Because generation is probabilistic, a regenerate is a genuinely new draw, not a retry of the same computation. Comparing two draws, or the same prompt on two models, is a fast quality check.

How models are made

Training

Training is the process of adjusting a model's parameters against data until its predictions get good.

It happens before you ever meet the model and costs the provider enormous compute. Nothing you type into a chat retrains the model in the moment, though providers may use conversations to train future versions depending on their policy.

Training data

Training data is the text and other material a model learned from.

Its breadth explains what a model knows, its gaps explain systematic blind spots, and its date range sets the knowledge cutoff.

Pretraining

Pretraining is the first, largest phase of training, where a model learns language and world knowledge from bulk text.

The result is capable but raw. Everything that makes a model behave like a helpful assistant comes afterward.

Fine-tuning

Fine-tuning is additional training on a narrower dataset to specialize or shape a pretrained model.

Providers fine-tune for helpfulness and safety; companies fine-tune for domain tasks. For most users, good prompting plus a capable general model beats paying for a fine-tune.

RLHF

RLHF, reinforcement learning from human feedback, is training a model against human ratings of its answers.

It is a large part of why modern assistants are polite, structured, and cautious. Overdone, it also produces the hedging that makes some models refuse or waffle.

Alignment

Alignment is the work of making a model's behavior match human intent and values.

In practice it covers everything from refusing harmful requests to simply following instructions accurately. When a model does something technically impressive but unwanted, that is an alignment gap.

Inference

Inference is running a trained model to produce an answer, as opposed to training it.

Every message you send triggers inference, and inference cost is what token pricing measures. Cheap-to-run models can be remarkably capable; the AI Model Cost Index shows a 2,000x price spread.

Transformer

The transformer is the neural network architecture behind essentially every modern language model.

Its key mechanism, attention, lets the model weigh every part of the input against every other part. You never need the details, but the word is everywhere in AI writing.

When things go wrong

Hallucination

A hallucination is a model stating something false with complete confidence.

It is not a bug that will be patched next release; it follows from how generation works. Frequency varies a lot by model and task. Treat any specific fact, number, or citation as unverified until checked. Why AI gets things wrong covers the mechanics.

Grounding

Grounding is giving a model authoritative material to answer from, instead of letting it rely on memory.

Pasting the contract, attaching the report, or enabling web search are all grounding. It is the most effective everyday defense against hallucination.

Bias

Bias is a model systematically skewing its outputs in ways inherited from training data.

It shows up as assumptions about people, places, and defaults. For consequential decisions, review AI output the way you would review a smart but unvetted intern's work.

Guardrails

Guardrails are the restrictions a provider builds around a model to block harmful or off-policy output.

Different providers draw the lines differently, which is a real product difference: the same request can be answered by one model and refused by another.

Jailbreak

A jailbreak is a prompt crafted to trick a model into ignoring its guardrails.

Providers patch them continuously. For ordinary users the term mostly matters as a security concept: anything a model can be talked into, someone will try to talk it into.

Red teaming

Red teaming is deliberately attacking a model before release to find harmful failure modes.

Providers run internal and external red teams, and published model cards often summarize what they found. It is the AI equivalent of penetration testing.

Knowledge cutoff

A knowledge cutoff is the date after which a model's training data ends.

Ask about anything more recent and the model either admits ignorance, browses the web if it can, or hallucinates. Knowing the cutoff of the model you are using prevents a whole category of errors.

Retrieval and your data

RAG (retrieval augmented generation)

RAG, retrieval augmented generation, is fetching relevant documents first and having the model answer from them.

It is how "chat with your knowledge base" products work under the hood, and why they can cite sources. RAG reduces hallucination but inherits the quality of whatever it retrieves.

Embedding

An embedding is a list of numbers representing the meaning of a piece of text.

Texts with similar meaning get nearby numbers, which lets software find related passages mathematically. Embeddings are the machinery behind semantic search and RAG.

Vector database

A vector database is a database built to store embeddings and find the nearest ones fast.

If a product says it "indexed" your documents for AI search, an embedding for every chunk is sitting in one of these.

Semantic search

Semantic search is finding text by meaning rather than by matching keywords.

Searching "money problems" can surface a passage about cash flow strain even though no word matches. It is embeddings put to work.

Chunking

Chunking is splitting long documents into pieces small enough to embed and retrieve well.

It sounds like plumbing and it is, but bad chunking is a common reason document-chat tools answer from the wrong part of a file.

Web search (in AI)

Web search in AI means the model fetching current pages before answering, instead of relying on training memory.

It trades speed for freshness and lets answers carry citations you can open. For anything after the model's knowledge cutoff, it is the difference between an answer and a guess.

Knowledge base

A knowledge base is the curated set of documents a company gives its AI to answer from.

Quality in, quality out: an assistant grounded on a stale wiki confidently serves stale answers, with citations.

Beyond text

Multimodal

Multimodal means a model that works across more than one medium, such as text plus images, audio, or video.

It is why you can screenshot an error and ask what is wrong. The multimodal AI guide covers what actually works today.

Vision model

A vision model is a model that can interpret images: photos, screenshots, charts, and documents.

Reading is not the same as generating; a model can describe your diagram perfectly and still be unable to draw one.

Image generation

Image generation is producing new images from a text description.

Quality, style range, and text rendering inside images vary sharply between generators, which is why serious users compare several. The image generator roundup stays current on the field.

Diffusion model

A diffusion model generates images by starting from noise and refining it step by step toward the description.

It is the dominant technique behind modern image generators. The word mostly matters so that "diffusion" in a product name tells you what the product does.

Speech to text

Speech to text is transcribing spoken audio into written words.

Modern transcription is accurate enough to make meetings, voice notes, and interviews searchable, and it is how voice input in chat apps works.

Text to speech

Text to speech is generating spoken audio from written words.

Current systems produce natural, expressive voices, which powers read-aloud features and AI voice conversations.

The ecosystem

API

An API is the programmatic doorway developers use to send prompts to a model from their own software.

API pricing is per token, which is where the per-answer economics of every AI product ultimately come from.

Open weights

Open weights means a model's trained parameters are published for anyone to download and run.

It enables self-hosting and lets other providers serve the model on their own infrastructure. DeepSeek and Llama are the famous examples.

Open-source model

An open-source model is, loosely, a model released for public use and modification, though licenses vary widely.

The practical questions are always the same two: can you use it commercially, and who is hosting it for you. The open weights ecosystem matters most when privacy or cost rules out hosted flagships.

Benchmark

A benchmark is a standard test used to score and compare models.

Useful for a first cut, routinely overfit in marketing. A model that tops a benchmark can still lose on your actual work, which is why running your own three-task comparison beats reading charts.

Agent

An agent is an AI system that takes actions in steps toward a goal, rather than just answering once.

Booking, researching, coding, and filing tickets are agent territory. The capability is real and improving fast; so is the need to review what an agent did before it matters.

Tool use

Tool use, or function calling, is a model invoking external capabilities like search, calculators, code, or your calendar.

It is how chat assistants act on the world instead of only describing it, and it is the mechanism underneath agents.

Reasoning model

A reasoning model is one that spends extra compute thinking through a problem internally before answering.

The trade is time and cost for accuracy on hard problems. For simple questions the extra thinking buys nothing, which is another argument for switching models per task.

Latency

Latency is how long you wait between sending a prompt and the answer arriving or starting to stream.

It varies by model size, load, and whether the model is a reasoning model. For interactive work, a fast good-enough model often beats a slow brilliant one.

Rate limit

A rate limit is the cap a provider puts on how many requests or tokens you can use in a time window.

Hitting one is the usual reason a chat product tells you to wait or upgrade. Message caps on subscription plans are rate limits with friendlier branding.

Keeping the glossary useful

Terms age fast in this field. This page is maintained as a reference: entries get rewritten when usage shifts, and the linked deep-dive guides carry the detail that does not fit a definition.

If a term you hit in the wild is missing here, the fastest route is to paste the sentence you found it in into a capable model and ask for the definition in context. In Whizi you can ask two models at once and notice when they disagree, which for new jargon is surprisingly often.