What are the basic AI terms everyone should know?
Every entry here opens with a one-sentence definition you can quote, then the detail. The 80 terms sit in groups by what you're doing, not by alphabet, so neighbours explain each other. The words move fast. Collins Dictionary picked "AI" as its word of the year for 2023 and "vibe coding" for 2025. If you only want the handful of words that matter in week one, AI terms explained is the gentler tour.
Artificial intelligence
Artificial intelligence is software that does tasks which normally need human judgment, such as writing, reasoning, or recognizing what's in an image.
John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon put the phrase into print in their Dartmouth workshop proposal of 31 August 1955. The EU AI Act defines an AI system as "a machine-based system that is designed to operate with varying levels of autonomy." In everyday talk in 2026, "AI" usually means a chat assistant built on a large language model.
Generative AI
Generative AI is AI that produces new content, such as text, images, audio, video or code, from a prompt.
ChatGPT, Claude, Gemini and every image generator are generative AI. Older AI mostly sorted or predicted things. This kind makes them.
AGI (artificial general intelligence)
AGI is a hypothetical AI that matches or beats people at most intellectual work, not just one task.
No one agrees on the test for it. Labs use the word as a goal, critics as a warning, and no shipping product is AGI.
Machine learning
Machine learning is the technique of teaching software by showing it examples instead of writing rules by hand.
Every modern AI assistant is built this way. Its behaviour comes from patterns in the training data, which is why one model can be brilliant on common tasks and shaky on rare ones.
Deep learning
Deep learning is machine learning done with neural networks that stack many layers.
It's the branch behind the current boom. Every large language model is a deep learning model.
Neural network
A neural network is a web of simple maths units, loosely modelled on neurons, whose connection strengths are learned from data.
Those learned strengths are the parameters further down this page. That's all the "neural" means.
Model
A model is one specific trained AI system, the thing you actually send a prompt to.
GPT, Claude and Gemini are model families. A family holds individual models with different sizes, speeds and prices, and picking 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 huge amounts of text to predict what text comes next.
That one trick, done at scale, produces answers, drafts, code and summaries. "Large" refers to both the training data and the parameter count.
NLP (natural language processing)
NLP is the field of getting computers to read, understand and produce human language.
It's the older umbrella term. Spam filters and translation apps were NLP long before chatbots, and LLMs are now its dominant tool.
Small language model
A small language model is an LLM cut down far enough to run cheaply, often on a phone or laptop.
Nobody agrees where "small" ends. You trade some quality for speed, privacy and price.
Chatbot
A chatbot is the conversational interface wrapped around a model, the thing with the message box.
The difference matters when you compare 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 chunk of text a model actually reads and writes, usually a word or a piece of one.
Pricing, speed and context windows are all counted in tokens. The labs don't define one the same way, though. Here's what each one's own docs said in November 2026, with the 1,000-word figure worked out where the docs give a words ratio:
| Lab | What its docs say a token is | 1,000 English words comes to about |
|---|---|---|
| OpenAI (API docs) | About 4 characters, or 0.75 words | 1,333 tokens |
| Google (Gemini docs) | About 4 characters; 100 tokens is 60 to 80 words | 1,250 to 1,667 tokens |
| Anthropic, Claude before 4.7 (glossary) | About 3.5 English characters | No words figure given; Anthropic points you to its token counting tool |
| Anthropic, Claude 4.7 and later (same page) | About 30 percent more tokens for the same text than earlier Claude models | About 30 percent more than the row above |
Tokenizer
A tokenizer is the fixed vocabulary and set of rules a model uses to cut text into tokens.
Each model family has its own. That's why the same paragraph costs a different number of tokens on Claude than on GPT.
Context window
A context window is the most text, counted in tokens, that a model can consider at once.
Your prompt, the chat so far, attached files and the answer all share it. In the AI Model Cost Index, 47 of the 100 priced models carry a window of a million tokens or more (prices fetched from OpenRouter on 20 August 2026). The context window comparison lists the current figure for every major model, and the context window explainer covers why it's behind most "the AI forgot" complaints.
Parameters
Parameters are the internal numbers a model learned during training, and the usual measure of its size.
They're counted in billions. OpenAI's open model gpt-oss-120b has about 117 billion, but only 5.1 billion of them work on any one token, per its August 2025 model card. More parameters loosely means more ability and more cost. As a shopping metric it's weak, and your own tests beat it.
Response
A response is the model's output for one prompt, also called a completion.
It's generated fresh every time. That's why regenerating the same prompt gives different wording and sometimes a different answer.
Which prompting terms change the answer you get?
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 can't see it. When a chatbot "has a personality", that's usually where it lives.
Memory
Memory is a chat app saving facts about you between conversations and feeding them back into later prompts.
The model itself doesn't remember anything. The app stores notes and pastes them in, so you can usually view and delete them.
Prompt engineering
Prompt engineering is the practice of writing prompts deliberately to get reliably better output.
Most of it is plain clarity: state the goal, give context, show an example, name the format. The prompt engineering guide for beginners covers the techniques that actually move results.
Context engineering
Context engineering is deciding everything a model sees on a turn, not just the prompt: instructions, memory, retrieved documents, tool results and chat history.
The term spread in 2025 as agents got longer jobs. A long-running agent fails more often from a cluttered context than from a badly worded request.
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 word mostly survives as a contrast with few-shot.
Few-shot
Few-shot means putting a handful of worked examples in the prompt so the model copies the pattern.
It's the cheapest reliability upgrade for format-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 it gives its final answer.
Asking for step-by-step reasoning measurably helps on hard problems. 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 it picks each next token.
Low temperature gives consistent, conservative output. High temperature gives variety and the odd bit of nonsense. Anthropic's own glossary warns that "even with temperature set to 0, the results will not be fully deterministic." Chat apps usually pick a middle value for you.
Streaming
Streaming is the model sending its answer token by token as it writes, instead of all at once at the end.
It's why answers seem to type themselves. You can also stop a bad answer early.
Regenerate
Regenerate means asking for a fresh response to the same prompt.
Generation is probabilistic, so a regenerate is a genuinely new draw, not a rerun of the same sum. Comparing two draws, or one prompt on two models, is a fast quality check.
How is an AI model trained, and what happens when you use it?
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 retrains the model on the spot, though a provider may use chats to train future versions, depending on its policy.
Supervised and unsupervised learning
Supervised learning trains on examples that come with the right answer attached, while unsupervised learning finds patterns in data that has no labels.
A spam filter trained on emails marked "spam" is supervised. An LLM learning from raw web text is closer to unsupervised. Researchers call it self-supervised, because the next word is the label.
Training data
Training data is the text and other material a model learned from.
Its breadth explains what a model knows. Its gaps explain the blind spots, and its date range sets the knowledge cutoff.
Overfitting
Overfitting is a model memorizing its training examples so closely that it does worse on anything new.
It's the classic machine learning failure. It's also why a model can ace a benchmark it accidentally saw in training and then stumble on your real task.
Pretraining
Pretraining is the first and largest phase of training, where a model learns language and world knowledge from bulk text.
The result is capable but raw. Everything that makes it act like a helpful assistant comes afterwards.
Fine-tuning
Fine-tuning is extra training on a narrower dataset to specialize or shape a pretrained model.
Providers fine-tune for helpfulness and safety, and companies fine-tune for domain tasks. For most people, good prompting on a capable general model beats paying for a fine-tune.
Distillation
Distillation is training a small model to copy the answers of a big one.
It's how labs ship cheap, fast versions that keep much of the big model's skill. Google said in May 2024 that Gemini 1.5 Flash was trained by 1.5 Pro this way.
RLHF
RLHF, reinforcement learning from human feedback, is training a model against people's ratings of its answers.
Anthropic's glossary describes the feedback as people ranking two or more example answers, with training nudging the model toward the higher-ranked ones. It's a big part of why assistants are polite and structured. Overdone, it also produces hedging and needless refusals.
Reinforcement learning
Reinforcement learning is training by trial and error, where a model gets a reward for good outcomes and learns to chase it.
RLHF is one use. Reasoning models are another: labs reward them for reaching correct answers on maths and code, and long step-by-step thinking emerges.
Alignment
Alignment is the work of making a model's behaviour match human intent and values.
In practice it covers everything from refusing harmful requests to simply following instructions accurately. When a model does something impressive but unwanted, that's 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 token pricing is what inference costs. Cheap-to-run models can be remarkably good, and the AI Model Cost Index shows a 2,000x price spread between the cheapest and priciest answer.
Quantization
Quantization is storing a model's parameters with fewer bits so it needs less memory and runs faster.
It costs a little accuracy. OpenAI ships gpt-oss-120b quantized to a 4-bit format so it fits within 80GB of memory, per its model card.
Transformer
The transformer is the neural network design behind essentially every modern language model.
A team at Google introduced it in the June 2017 paper "Attention Is All You Need", which opens: "We propose a new simple network architecture, the Transformer, based solely on attention mechanisms." Attention lets the model weigh every part of the input against every other part. That's the T in GPT.
Mixture of experts
A mixture of experts is a model split into many sub-networks, where only a few switch on for each token.
You get the knowledge of a huge model at the running cost of a much smaller one. gpt-oss-120b is one, and so are the DeepSeek V3 family models.
What do hallucination, bias and jailbreak mean in AI?
Hallucination
A hallucination is a model stating something false with complete confidence.
It isn't a bug the next release will patch. It follows from how generation works, and the rate varies a lot by model and task. Treat any specific fact, number or citation as unchecked until you've checked it. Why AI gets things wrong covers the mechanics.
Grounding
Grounding is giving a model trusted material to answer from, so it doesn't lean on memory.
Pasting the contract, attaching the report and turning on web search are all grounding. It's the best everyday defence against hallucination.
Bias
Bias is a model skewing its outputs in a consistent direction it picked up from training data.
It shows up as assumptions about people, places and defaults. For decisions that matter, review AI output the way you'd review a smart but untested intern.
Sycophancy
Sycophancy is a model telling you what you want to hear instead of what's true.
It agrees with a wrong premise, praises a weak draft, or drops a correct answer when you push back. Ask it to argue the other side and you'll often see it.
Guardrails
Guardrails are the limits a provider builds around a model to block harmful or off-policy output.
Providers draw the lines in different places. That's 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 all the time. For ordinary users it's mainly a security idea: anything a model can be talked into, someone will try.
Prompt injection
Prompt injection is hidden instructions planted in content a model reads, such as a web page, email or file, that hijack what it does.
It's the jailbreak's more dangerous cousin. You don't type it. An agent reading a booby-trapped page can be told to leak data or click the wrong thing.
Red teaming
Red teaming is deliberately attacking a model before release to find harmful failure modes.
Providers run internal and outside red teams. It's the AI version of penetration testing.
Model card
A model card is the document a lab publishes with a model, listing what it's for, how it was tested and where it fails.
It's where red-team findings, benchmark scores and known limits get written down. OpenAI's gpt-oss model card is a typical one.
Knowledge cutoff
A knowledge cutoff is the date after which a model's training data ends.
Ask about anything later and the model admits it doesn't know, searches the web if it can, or makes something up. Knowing your model's cutoff prevents a whole class of errors.
What do RAG, embeddings and vector databases mean?
RAG (retrieval augmented generation)
RAG, retrieval augmented generation, is fetching relevant documents first and having the model answer from them.
It's how "chat with your knowledge base" products work, and why they can cite sources. Anthropic's glossary puts the catch plainly: RAG's effectiveness "depends on the quality and relevance of the external knowledge base."
Embedding
An embedding is a list of numbers that stands for the meaning of a piece of text.
Texts with similar meanings get nearby numbers, so software can find related passages with maths. 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 sits in one of these.
Semantic search
Semantic search is finding text by meaning instead of by matching keywords.
Searching "money problems" can surface a passage about cash flow strain with no word in common. It's embeddings at work.
Chunking
Chunking is splitting long documents into pieces small enough to embed and retrieve well.
It sounds like plumbing, and it is. Bad chunking is still a common reason a document chat answers from the wrong part of a file.
Web search (in AI)
Web search in AI is the model fetching current pages before it answers, instead of answering from training memory.
It trades speed for freshness and lets answers carry links you can open. For anything after the knowledge cutoff, it's the difference between an answer and a guess.
Deep research
Deep research is an agent mode that runs dozens of searches, reads the results and writes a cited report.
OpenAI, Google and Perplexity all ship a version. A run takes minutes, not seconds, and the citations still need checking.
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 serves stale answers, with citations.
Three terms in this glossary are rival answers to one question, how to make a model better at your task, and mixing them up is the most common jargon mistake in product meetings:
| Approach | What changes | Cost and speed | When it wins |
|---|---|---|---|
| Better prompting (zero-shot, few-shot) | Only the text you send | Free, takes minutes | Almost always the first move; two or three worked examples fix most format problems |
| RAG | What the model reads before answering | Moderate to build, no retraining | Answers that must stay current or cite your own documents |
| Fine-tuning | The model's own parameters | Slow and expensive, then fixed | Stable style and format needs at scale, almost never for one person |
Which AI terms cover images, voice and video?
Multimodal
Multimodal means a model that works across more than one medium, such as text plus images, audio or video.
It's why you can screenshot an error and ask what's 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 isn't drawing. A model can describe your diagram perfectly and still be unable to make one.
Computer vision
Computer vision is the wider field of getting software to understand images and video.
Face unlock, number-plate cameras and photo search all run on it. Vision models are its newest, chattiest branch.
Image generation
Image generation is producing new images from a text description.
Quality, style range and text inside images vary sharply between generators, so serious users compare several. The best AI image generators roundup keeps up with the field.
Deepfake
A deepfake is AI-made image, audio or video that shows a real person saying or doing something they never did.
The tech is the same as ordinary image and voice generation. The word marks the misuse.
Diffusion model
A diffusion model makes images by starting from random noise and refining it step by step toward the description.
It's the main technique behind modern image and video generators. When "diffusion" shows up in a product name, now you know what it does.
Speech to text
Speech to text is turning spoken audio into written words.
It's accurate enough now to make meetings, voice notes and interviews searchable. It's also 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. They power read-aloud features and AI voice chats.
What are AI agents, MCP and the other newer AI terms?
API
An API is the programmatic doorway developers use to send prompts to a model from their own software.
API pricing is per token. That's where the per-answer economics of every AI product come from.
Open weights
Open weights means a model's trained parameters are published for anyone to download and run.
It lets you self-host and lets other providers serve the model. DeepSeek, Llama and Qwen are the famous families. OpenAI released gpt-oss on 5 August 2025 under the Apache 2.0 licence, its first open-weight language models since GPT-2.
Open-source model
An open-source model is, loosely, a model released for public use and changes, though licences vary widely.
Two practical questions always follow: can you use it commercially, and who's hosting it for you? How to choose an AI model covers when open models beat hosted flagships on privacy or cost.
Benchmark
A benchmark is a standard test used to score and compare models.
It's fine for a first cut and routinely overfit in marketing. A model that tops a chart can still lose on your actual work, so your own three-task comparison beats reading charts.
Turing test
The Turing test asks whether a person chatting by text can tell a machine from a human.
Alan Turing proposed it as "the imitation game" in his 1950 paper "Computing Machinery and Intelligence." In a March 2025 study at UC San Diego, people picked GPT-4.5 as the human 73% of the time. Passing it doesn't prove intelligence.
Agent
An agent is an AI system that takes actions in steps toward a goal, not just a single reply.
Booking, researching, coding and filing tickets are agent work. The ability is real and improving fast. So is the need to check what an agent did before it matters.
Agentic AI
Agentic AI is the umbrella term for products built around agents that plan, use tools and work for minutes or hours with little supervision.
Read it as "this one acts", then ask what it's allowed to touch.
Vibe coding
Vibe coding is building software by describing what you want to an AI and accepting the code it writes, often without reading it closely.
Andrej Karpathy coined it in February 2025, and Collins named it word of the year that November. It's great for prototypes. It's risky for anything holding real data.
Tool use
Tool use, or function calling, is a model invoking outside abilities like search, calculators, code or your calendar.
It's how chat assistants act on the world instead of only describing it. It's also the mechanism underneath agents.
MCP (Model Context Protocol)
MCP is an open standard for plugging tools and data sources into any AI app, so one connector works everywhere.
Anthropic published it on 25 November 2024. OpenAI and Google adopted it within months, and on 9 December 2025 Anthropic handed it to the Linux Foundation's new Agentic AI Foundation, which counted more than 10,000 MCP servers at launch.
Reasoning model
A reasoning model is one that spends extra compute thinking a problem through internally before it answers.
The trade is time and cost for accuracy on hard problems. For simple questions the extra thinking buys nothing, which is one more reason to switch models per task.
Latency
Latency is how long you wait between sending a prompt and the answer arriving or starting to stream.
It depends on model size, load, and whether the model reasons first. 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 app tells you to wait or upgrade. Message caps on subscription plans are rate limits with friendlier branding.
What should you do when a term is not in this glossary?
Paste the sentence you found the term in into a capable model and ask for the definition in that context. New AI jargon is often coined in one product launch and means something slightly different in the next, so context beats a dictionary. This page gets rewritten when usage shifts, and the linked guides carry the detail a definition can't.
Ask two models, not one. On Whizi Pro, side by side model chats put GPT, Claude or Gemini on the same prompt, and for brand-new jargon they disagree more often than you'd expect. When they do, the newer usage usually wins.
- Quote the one-sentence definition, then read the detail before relying on it
- When two sources define a term differently, the newer usage usually wins in AI
- Learn token, context window and hallucination first; most other terms hang off those three
- Count tokens with the tokenizer of the model you'll use: by OpenAI's and Google's own figures, 1,000 words runs from 1,250 to 1,667 tokens, and Claude 4.7 and later count about 30 percent more than earlier Claude
- Use the beginner glossary first if this page feels like a reference manual, because it is one
Frequently asked questions
What is the difference between AI, machine learning, and an LLM?
AI is the broad goal of software doing tasks that need human judgment. Machine learning is the technique of teaching software from examples, and it's how nearly all modern AI is built. An LLM is one kind of machine-learned model, trained on text, and it's the kind behind ChatGPT, Claude and Gemini.
What does RAG mean in AI?
RAG stands for retrieval augmented generation: the system first fetches relevant documents, then has the model answer from them instead of from memory. It's the standard technique behind products that chat with your files or knowledge base, and it's why those products can cite sources.
What is the difference between an AI agent and a chatbot?
A chatbot answers you, and an agent acts for you. An agent plans steps, calls tools such as search, code or your calendar, and keeps going until the goal is done. ChatGPT, for one, is both: a chatbot by default, with an agent mode you switch on.
What is a token in AI?
A token is the unit of text a model reads and writes, usually a word or part of one. OpenAI's rule of thumb is about 4 characters or 0.75 words per token, so 1,000 English words is about 1,333 tokens. Google's Gemini docs give 60 to 80 words per 100 tokens, and Anthropic says Claude 4.7 and later produce about 30 percent more tokens than earlier Claude models.
Which AI terms should a beginner learn first?
Learn five first: model, prompt, token, context window and hallucination. They explain most of what you'll see in a chat app, from pricing to "the AI forgot". The beginner version of this glossary covers those with more patience, and the rest of this page is for when an article throws a term at you.