What is a token in AI?

A token is the chunk of text a model actually processes — usually part of a word rather than a whole one. In English, a rough working figure is about four characters, or three-quarters of a word, per token. It matters because pricing, rate limits and context windows are all measured in tokens, not words, and the same text costs more tokens in some languages than in English.

Retrieval-augmented generation. Rather than relying on what a model absorbed during training, the system first retrieves relevant documents and places them in the prompt, so the answer is generated from material you control. It is how most tools answer questions about your own files, and it is usually a better fix for wrong answers than fine-tuning.

What is the difference between a jailbreak and a prompt injection?

A jailbreak is a user trying to talk a model out of its own rules. A prompt injection is instructions hidden inside content the model reads — a web page, a PDF, an email — which the model then follows as though you had typed them. The distinction matters because injection is an attack on you by a third party, not by you on the model, and it is the more serious security problem.

What is a context window?

The maximum amount of text, measured in tokens, that a model can consider at once — the prompt, any attached documents, the conversation so far, and the reply it is generating, all counted together. When a conversation exceeds it, the earliest material stops influencing the answer, which is why long chats appear to forget things.

A model given tools and a goal, allowed to take multiple steps on its own — searching, running code, editing files, calling APIs — and to decide what to do next based on what happened. The difference from a chatbot is not intelligence but autonomy: an agent acts between your instruction and its answer.

Do I need to know these terms to use AI tools?

Not most of them. About ten do real work in everyday use: token, context window, prompt, system prompt, hallucination, RAG, agent, open weights, rate limit and prompt injection. The rest are useful when you are comparing products or reading documentation, which is exactly when vendors rely on you not knowing them.

AI Glossary (2026): The Terms That Actually Matter

Q: What is the difference between open source and open weights?

Open weights means you can download the model's learned parameters and run it yourself. Open source, in the sense the term normally carries, would also mean the training code and data are available and the licence permits any use. Most models marketed as open are open-weights only, often under licences that restrict commercial use or scale. The distinction matters legally, not just semantically.

Most AI glossaries define a hundred terms as though each were equally load-bearing. They are not. A handful genuinely change what you do; some are precise technical terms worth knowing exactly; and a few are marketing that survives because nobody asks what it means.

This one is grouped by what you would be doing when you meet the word, and says plainly when a term is doing less work than it appears to.

The models themselves

Large language model (LLM) A model trained on very large amounts of text to predict what comes next. Everything a chatbot does — answering, summarising, translating, writing code — is that one operation applied repeatedly. Worth remembering when output is confidently wrong: the model is not consulting knowledge, it is continuing a pattern.

Parameters The learned numbers inside a model, adjusted during training. Counts are quoted in billions. More parameters once reliably meant a better model; it no longer does, because training data quality, architecture and post-training now matter as much. Treat parameter counts as trivia rather than a spec.

Token The unit of text a model actually reads and writes — typically a fragment of a word. A rough English heuristic is four characters, or three-quarters of a word, per token. Pricing, rate limits and context windows are all denominated in tokens, and non-English text usually consumes more of them for the same meaning.

Context window The maximum tokens a model can hold in view at once: your prompt, attached files, the conversation history, and the answer being generated, all counted together. Exceeding it does not produce an error — the oldest material simply stops influencing the reply, which is why long conversations seem to develop amnesia.

Foundation model / frontier model A large general-purpose model that other products are built on. “Frontier” informally means the most capable current generation. Neither term has a technical definition, and “frontier” in particular is a marketing position as often as a description.

Multimodal Handles more than text — images, audio, video, or some combination — natively rather than by handing off to a separate system. Check which directions are supported: many tools read images but cannot produce them.

Reasoning model A model trained to spend additional computation working through a problem before answering, rather than replying immediately. Better at maths, logic and multi-step code; slower and dearer per answer. Not better at everything, and rarely worth it for straightforward writing.

Mixture of experts (MoE) An architecture that routes each token through a small subset of the network rather than all of it, so a model can be very large in total while only a fraction runs per token. It is why some models advertise huge parameter counts at modest speed.

Inference Running a trained model to get an answer, as opposed to training it. Your monthly subscription buys inference. “Inference cost” is what vendors are managing when free tiers get throttled.

Fine-tuning Additional training on a narrower dataset to adapt a model’s behaviour or style. Usually the wrong first answer to “the model doesn’t know about my business” — retrieval (RAG, below) is cheaper, faster to update, and easier to audit.

Distillation Training a smaller model to imitate a larger one. It is how cheap fast models with surprisingly good behaviour are made, and it is why capability propagates across vendors quickly.

Open weights vs open source Not synonyms, and the difference is legal rather than pedantic. Open weights means the trained parameters can be downloaded and run. Open source, used properly, also implies the training code and data are available and the licence permits any use, including commercial. Most “open” models are open-weights only, under bespoke licences with restrictions. Read the licence before you build on one.

Temperature A setting controlling randomness. Low values make output more predictable and repetitive; high values more varied and more likely to go wrong. Exposed in APIs and developer tools, usually hidden in consumer apps.

Working with them

Prompt What you send the model. Everything downstream depends on it, which is why prompt engineering became a discipline and then, mostly, became ordinary writing skill.

System prompt Standing instructions that sit above the conversation and shape every reply — role, tone, constraints, output format. In consumer products it is set by the vendor and usually invisible to you; in developer tools it is yours to write, and it is the highest-leverage text in the whole system.

Zero-shot / few-shot Zero-shot is asking with no examples. Few-shot is including two or three worked examples of what you want. Few-shot remains the single most reliable way to fix output that is nearly right but in the wrong shape.

Chain of thought Prompting a model to work through intermediate steps before committing to an answer. Reasoning models now do this internally, so explicitly asking for it matters less than it did — but it still helps on older or smaller models.

Retrieval-augmented generation (RAG) Retrieving relevant documents and putting them into the prompt so the answer is generated from material you control rather than from training data. It is what NotebookLM does over your uploads and what most “chat with your documents” features are underneath. See the research guide for when this matters.

Grounding Tying an answer to identifiable sources — retrieved documents, search results, a database. A grounded answer can still be wrong, but it can at least be checked, which ungrounded output cannot.

Hallucination Output that is fluent, confident and false. The term is criticised for implying perception the model does not have; the useful thing to understand is that a model has no internal signal distinguishing a fact it absorbed from a plausible continuation it invented, so both arrive in the same untroubled tone.

Agent A model given tools and a goal, permitted to take several steps by itself and to decide the next one from what happened. The distinction from a chatbot is autonomy, not intelligence. In coding this is the whole assistant-versus-agent split covered in the AI coding guide.

Tool use / function calling The mechanism behind agents: the model is told which functions exist and emits a structured request to call one, which your system executes and returns. It is how a language model books a flight or reads a file — it never does either itself.

Model Context Protocol (MCP) An open standard for connecting AI assistants to external tools and data through one interface, so a capability written once works across any client that speaks it. Increasingly the default answer to “how do I let an assistant use my system?”

Guardrails Constraints on what a model will produce — training, filters on input and output, or both. Worth knowing they are probabilistic rather than absolute.

Jailbreak A user deliberately working around a model’s restrictions.

Prompt injection Instructions hidden inside content the model consumes — a web page, a PDF, an email, a code comment — which the model then follows as if you had written them. This is the one to actually worry about: it is a third party attacking you through the model, it has no complete fix, and it gets more dangerous the more tools an agent can reach.

Images, video and audio

Diffusion model Generates an image by starting from random noise and repeatedly removing it, steered toward the prompt. Behind most image tools including Midjourney, FLUX and Ideogram.

Seed The number that fixes the random starting point. Same prompt plus same seed gives the same image, which is what makes iteration possible rather than a slot machine.

Inpainting / outpainting Regenerating a selected region of an existing image, or extending it beyond its original edges.

Upscaling Increasing resolution, inventing plausible detail as it goes. Fine for print size; not evidence of anything, since the added detail was never in the original.

Voice cloning Reproducing a specific voice from a sample. Legally and ethically the sharpest edge in generative audio — consent from the voice’s owner is the requirement, and ElevenLabs and its peers gate this behind verification for that reason.

Text-to-video Generating video from a written description, usually in short clips. Cost is metered in credits rather than minutes, which hides the real price — see what AI video credits actually buy.

Content Credentials / C2PA A provenance standard that attaches tamper-evident metadata recording how a file was made and edited. Adopted by several major tools. It marks content that carries it; it cannot mark content that does not, so absence proves nothing.

Deepfake Synthetic media depicting a real person doing or saying something they did not. The term covers both the technique and the deception; the technique underlies legitimate avatar video products too.

Buying and running them

Freemium A free tier with real limits and a paid tier that lifts them. The interesting question is always which specific capability the free tier meters — in coding tools it is almost always agent work rather than autocomplete.

Credits A vendor’s internal currency, priced so that comparing across vendors is difficult. Always convert to what you actually get: minutes of video, images, or agent tasks.

Rate limit A cap on how much you can use in a window. The detail that matters is the window: a five-hour rolling limit behaves very differently from a monthly quota, and vendors often decline to state which they use.

Seat One user’s licence. Team pricing is usually per seat per month, which is where apparently cheap tools become expensive.

API vs app The app is the product you log into; the API is programmatic access, billed per token rather than per month. Same underlying model, different pricing, different data handling terms.

Zero data retention A contractual arrangement, usually API-only and often enterprise-only, under which prompts and outputs are not stored after the request. Distinct from “we don’t train on your data” — which addresses training, not storage.

Data residency Which country your data is processed and stored in. Relevant to GDPR and to sector-specific rules, and frequently the reason an otherwise ideal tool is unusable at an organisation.

Benchmark A standardised test of model capability. Useful for coarse ranking, weak evidence for your specific work, and vulnerable to contamination when test material has leaked into training data. Vendors quote the benchmarks they win.

The ones that mean less than they sound

AI-powered — an unqualified claim, applied to everything from a frontier model to a sorting rule. Ask which model, doing what.

Agentic — often just “the software does more than one thing before stopping.” Meaningful when it describes tool use and autonomy; decorative otherwise.

Hallucination-free — no current system can promise this. Grounding reduces the rate; it does not eliminate the failure mode.

Trained on your data — ambiguous by design. It can mean fine-tuning, or retrieval at query time, or nothing beyond your files being uploaded. Ask which.

Human-level — undefined without saying which humans at which task.

The ten worth knowing today

If you are only going to keep ten: token, context window, prompt, system prompt, hallucination, RAG, agent, open weights, rate limit and prompt injection. Those cover almost every practical decision about what to use, what it will cost, and what can go wrong.

The rest earn their place when you start comparing products closely — which is precisely the moment the vocabulary stops being neutral. Our reviews and comparisons translate the marketing back into these terms wherever we can.