The plain-English AI glossary: every term Australians actually need

Pricing figures updated 6 September 2026: USD prices are converted at the current rate (1 USD ≈ 1.40 AUD, verified 5 September 2026) and shown ex-GST, replacing the retired 1.55 + GST basis. A GST-registered Australian business that supplies its ABN is not charged GST on an imported digital service; other buyers should allow ~10% on top (Methodology).

By Adam, Editor

Last reviewed: 14 June 2026

Verdict (TL;DR): You don't need to understand how AI works to use it well — but you do need to understand the words, because every tool, sales page and review leans on them. This is the decoder. Roughly forty terms that genuinely come up when an Australian small business or professional looks at AI, each explained in one plain sentence, with a second line on why it matters to you — not a lecture. Skim it once, then keep it open in a tab. If you only learn four, learn these: a hallucination is when the AI states something false with total confidence (so always check); the free tier is usually enough to start (so don't pay early); data residency is about where your information physically sits (which matters in Australia); and an agent is just an AI that does steps, not only writes about them. The rest is detail.

How to use this page

The terms are grouped, not alphabetical, because that's how they come up in real life — the basics first, then how the tools behave, then money, then the Australia-and-safety words, then the buzzwords you can mostly ignore. Use your browser's Find (Ctrl-F or ⌘-F) to jump to a word.

A companion note: if you're brand new, start with our first-month plan for getting going with AI, then come back here whenever a word trips you up.

The basics — what the things actually are

Artificial intelligence (AI). Software that does tasks we used to think needed a human — writing, summarising, answering questions, recognising images. In 2026 the term mostly means the chat-style tools below, not robots. In practice: when someone says "use AI," they almost always mean a tool like ChatGPT, Claude or Gemini.

Large language model (LLM). The engine inside a chat assistant — a system trained on enormous amounts of text that predicts what words should come next. That prediction is good enough to feel like understanding. In practice: you never touch the model directly; you use a product built on top of it. ChatGPT is the product; GPT is the model underneath.

Model. A single trained version of that engine, usually with a name and a number (the numbers go up as they improve). One product often offers several models — a fast cheap one and a slower smarter one. In practice: the "which model" dropdown matters less than people think. Start on the default; switch to the smarter model only when an answer disappoints.

Generative AI. AI that creates something new — text, an image, audio, video, code — rather than just sorting or classifying what already exists. In practice: this is the whole current wave. An image generator and a writing assistant are both generative AI.

Chatbot / assistant. A tool you talk to in plain language, in a back-and-forth. "Assistant" usually implies it can do more than chat — draft, summarise, search, sometimes act. In practice: the general assistants are the Swiss-army knife; the specialised tools below do one job better. Most people need one of each.

Prompt. What you type in — your instruction or question. The single biggest lever on output quality is a clear prompt. In practice: "fix this" gets you mush; "shorten this email to three sentences, keep it polite, no exclamation marks" gets you something usable. Our guide on getting better answers is entirely about this.

Output / response / completion. What the AI gives back. "Completion" is the technical word; "output" is fine. In practice: treat every output as a strong first draft, never a finished product.

How the tools behave — the words that explain the quirks

Hallucination. When the AI states something false — a fake statistic, a court case that never happened, a made-up quote — with complete confidence. It isn't lying; it's predicting plausible words, and sometimes plausible is wrong. In practice: this is the one to burn into memory. Never publish, send or rely on a fact, figure, name or quote from an AI without checking it against a real source. The risk is highest in law, finance and health.

Token. The small chunk of text a model reads and writes in — roughly three-quarters of a word in English. Tools measure usage and limits in tokens, not words. In practice: you rarely need to think about tokens, except that "1,000 tokens" is about 750 words, and longer conversations cost more of them.

Context window. How much the model can "hold in its head" at once — your prompt plus everything earlier in the conversation plus any document you've pasted. Measured in tokens. Go past it and the tool starts forgetting the beginning. In practice: if a long chat starts contradicting itself or losing the thread, the context window is full. Start a fresh chat and paste in just what matters.

Temperature. A setting (mostly hidden in consumer tools) that controls how predictable versus creative the output is. Low temperature = safe and repetitive; high = surprising and sometimes off the rails. In practice: you'll rarely see this dial, but it's why the same prompt can give a slightly different answer each time.

Training data. The text, images and code a model learned from before you ever used it. It has a cut-off date, so the model doesn't natively know about events after that point. In practice: this is why an assistant may not know last week's news unless it can search the web. Ask "what's your knowledge cut-off?" if it matters.

Knowledge cut-off. The date after which the model wasn't trained on anything. Anything more recent, it either doesn't know or has to look up. In practice: for current prices, news or law, prefer a tool that searches the live web (Perplexity is built around this) over one relying on training alone.

Grounding / citations. When a tool backs its answer with real, linkable sources instead of pulling from memory. "Grounded" answers are checkable; ungrounded ones are just assertions. In practice: prefer tools that show their sources, especially for anything you'll act on. A citation you can click is worth ten confident sentences you can't.

Multimodal. A tool that handles more than text — it can read an image you upload, listen to audio, or look at a PDF, not just words. In practice: "multimodal" is why you can photograph a handwritten invoice and ask the AI to type it up.

Reasoning model. A newer class of model that "thinks" for longer before answering — working through steps internally — and so does better on maths, logic and multi-step problems, at the cost of speed. In practice: use the reasoning option for a tricky analysis or a problem with several moving parts; use the fast default for everyday drafting.

Money — the words on the pricing page

Free tier. A genuinely usable version that costs nothing, usually with limits (slower, fewer messages, older model). Nearly every major tool has one. In practice: start here, always. Most people get real value without ever paying. Don't upgrade until a free tool has clearly earned it — that's the whole point of our five questions before paying guide.

Freemium. The business model behind the free tier: free to start, pay to unlock more. The free version is real, not a trial — it just has a ceiling. In practice: the ceiling is where they want you to upgrade. Notice whether you're actually hitting it or just being nudged.

Trial. Time-limited full access that expires (often 7 or 14 days), sometimes needing a card up front. Different from a free tier, which doesn't expire. In practice: if a card is required, set a phone reminder to cancel before it bills.

Subscription / per-seat. A recurring monthly or yearly fee, often charged "per seat" — i.e. per person who uses it. Five staff on a per-seat tool means five times the price. In practice: for a small team, per-seat pricing adds up fast. Check whether you need everyone on it or just one or two.

Credits / usage-based. Instead of (or on top of) a flat fee, you spend credits per action — per image generated, per minute transcribed, per word written. Run out, top up. In practice: easy to overspend without noticing. If a tool is credit-based, watch the meter for the first month.

API. A way for software to talk to the AI directly, used by developers to build the AI into other apps. Priced by usage, not a subscription. In practice: you almost certainly don't need the API. If a tool offers "API access" as a feature, it's not aimed at you unless you have a developer.

Enterprise / quote-only. The top tier, aimed at big organisations, usually with no public price — you have to "contact sales." Adds admin controls, security guarantees and a contract. In practice: "contact sales" is a fair sign a tool isn't built for a sole trader or small team. Plenty of excellent tools never make you ring anyone.

GST. Australia's 10% Goods and Services Tax. Many AI tools price in US dollars and either add GST at checkout or bake it in, so the figure you finally pay is higher than the sticker. In practice: a "US$20/month" tool lands near A$28/month once you convert and add GST. We show the ex-GST Australian figure in our reviews so there's no nasty surprise — see how we score and price.

Australia and safety — the words that matter most locally

Data residency. Where your information physically lives — which country's servers it sits on. It's the defining question for Australian businesses handling client or customer data. In practice: "stored in Australia" and "processed overseas" can both be true of the same tool. Our full guide to AI data residency unpacks why that distinction is the whole game.

At-rest vs inference (storage vs processing). Two different things vendors blur together. At-rest is where your data is stored; inference is where it's sent to be processed when the AI answers. A tool can store your files in Sydney yet still send each prompt overseas to be answered. In practice: if data residency matters to you, ask both questions separately. A local storage promise alone doesn't mean your prompts stay onshore.

Training opt-out. A setting (or a default) that stops the vendor from using your conversations to train future models. Business and paid tiers often switch this off by default; free tiers sometimes don't. In practice: if you paste anything sensitive, find this setting first. The rule of thumb still holds: don't paste what you wouldn't email to a stranger.

On-device / local AI. AI that runs entirely on your own computer or phone, so your data never leaves it. Slower and less capable than the big cloud tools, but private by design. In practice: worth it only when privacy is non-negotiable. See our round-up of local and offline AI for who actually needs it.

Privacy Act / APPs. Australia's privacy law and its Australian Privacy Principles, which govern how businesses handle personal information — including what you feed into an AI tool. In practice: you're still responsible for client data after it goes into an AI. The tool's compliance doesn't transfer your obligations away. This is general information, not legal advice — check with a professional for your situation.

Sovereign / onshore. Marketing words for "kept in Australia." "Sovereign AI" usually means hosted on Australian infrastructure under Australian control. In practice: treat the label as a claim to verify, not a guarantee. Ask exactly what's onshore — storage, processing, support, or all three.

IRAP. An Australian government security-assessment framework. If a tool is "IRAP assessed," it's been checked against standards used by government and regulated industries. In practice: only relevant if you work with government or in a regulated sector. For a café or a marketing agency, its absence isn't a dealbreaker.

The buzzwords — decoded, then mostly ignorable

Agent / agentic AI. AI that doesn't just answer — it does: takes a goal, plans steps, and carries them out across apps (books the thing, fills the form, sends the email). The big shift of 2026. In practice: genuinely useful and genuinely riskier, because it acts on your behalf. Start it on low-stakes tasks and keep a human checking the result.

RAG (retrieval-augmented generation). A method where the AI looks something up in a trusted source — your documents, a knowledge base — before answering, so it quotes real material instead of guessing. In practice: you don't need the acronym; you need the benefit. It's how "chat with your own documents" features avoid making things up.

Fine-tuning. Taking a general model and training it further on specific material so it specialises — a model tuned on legal contracts, say. In practice: a developer concept. As a user you'll rarely do it; you'll just use a tool someone else fine-tuned.

Prompt engineering. The (slightly grand) name for writing good prompts. It's a real skill but not a mysterious one — clarity, context and an example will get you 90% of the way. In practice: ignore the courses promising to make you a "prompt engineer." Our getting-started plan covers what you actually need.

Open source / open weights. Models anyone can download and run themselves, free, rather than only renting through a company's app. In practice: matters if you want local, private AI (see above). For most people the polished paid apps are the easier path.

Parameters. A rough measure of a model's size, quoted in billions. Bigger isn't automatically better, and the number tells you almost nothing about whether a tool suits you. In practice: safely ignore. Judge a tool by what it does for your job, not its spec sheet.

Benchmark. A standardised test used to score models against each other. Useful for researchers, routinely cherry-picked in marketing. In practice: "tops the benchmarks" is a sales line. Your own test on your own task is the only benchmark that counts.

AGI (artificial general intelligence). A hypothetical future AI as broadly capable as a human across everything. It does not exist, and the term is used loosely (often to sell things). In practice: irrelevant to choosing a tool today. Don't let it into a buying decision.

The shortest version

If this page is too much, here's the whole thing in six lines:

• A model is the engine; the app is what you actually use.
• A hallucination is a confident wrong answer — always check facts, names and figures.
• The free tier is usually enough; don't pay until it's earned.
• Data residency is about where your data physically sits — storage and processing are separate questions.
• An agent acts, not just writes — useful, but keep a human watching.
• Most of the rest is marketing. Judge a tool by whether it saves you time on a real job.

Frequently asked questions

What's the difference between ChatGPT and an LLM? The LLM (a model like GPT) is the engine. ChatGPT is the product built around it — the app, the chat window, the extras. You use the product; the model works underneath.

What does "hallucination" mean in AI? It's when the tool states something false — a made-up fact, figure, quote or source — with full confidence. It isn't deliberate; the model predicts plausible words and sometimes plausible is wrong. Always verify anything factual before you rely on it.

Do I need to understand how AI works to use it? No. You need to understand the words on the screen and a few safety habits — check facts, mind your data, start with the free tier. The mechanics underneath are optional.

What's a "token" and why does it matter? A token is a small chunk of text — about three-quarters of a word. Tools measure limits and usage in tokens. You rarely need to think about it, except that longer chats and bigger documents use more of them.

What's the difference between data storage and data processing for AI? Storage ("at rest") is where your data sits long-term; processing ("inference") is where it's sent to be worked on when the AI answers. A tool can store your data in Australia yet still process it overseas — so ask both questions, not one. Our data residency guide covers it in full.

Is an "AI agent" the same as a chatbot? No. A chatbot answers; an agent acts — it carries out multi-step tasks across apps on your behalf. Agents are more useful and more risky, so keep a human checking what they do.

Sources

This is a reference glossary of established, widely-used terminology in the AI field; the definitions are explanatory and not tied to any single vendor's claim. Tool examples link to our own verified reviews and round-ups, each of which carries its own primary sourcing. Australia-specific points (GST treatment, data residency, the Privacy Act and Australian Privacy Principles, IRAP) are explained at a general level — for the detail behind the residency points, see our AI data residency guide, which synthesises separately-sourced reporting.

This glossary is general information, not legal, financial or technical advice. For decisions that turn on privacy, compliance or contractual obligations, consult a qualified professional.

Part of The AI Directory's beginner series: where to start · five questions before paying · data residency · this glossary. Reviewed and maintained by Adam, Editor.