The Plain-English AI Glossary: 25 Terms That Actually Matter in 2026

There is a specific moment that stops people from using AI, and it usually isn't the technology.

It's a sentence like this one, pulled from a product announcement: "Our new multimodal foundation model supports a 200K context window with improved RLHF alignment and reduced hallucination rates."

If you're a marketing director, a small business owner, or someone who has run a career perfectly well for thirty years without needing to know what a "context window" is, that sentence does one thing: it tells you this isn't for you.

It is for you. That sentence just says the software can look at pictures as well as text, can hold about a novel's worth of your conversation in mind at once, has been trained to be more helpful, and makes things up slightly less often than it used to.

That's the whole translation. Every term in it is ordinary once someone bothers to explain it in ordinary words.

This is a guide to the twenty-five terms that actually come up — not the ones that show up in research papers, but the ones you'll hit reading a product page, an article, or a settings menu. Each one gets a plain definition and, where it matters, a note about why you'd care.

The full glossary runs to 91 terms and lives at the AI Foresights AI Glossary. This piece covers the ones worth knowing first.

---

Start here: the four terms everything else sits on

Artificial intelligence (AI)

Software that does things we used to think required a person — recognizing a face, writing a paragraph, answering a question in a way that makes sense.

That's it. There's no precise technical line where a program becomes "AI." The term has stretched to cover everything from your spam filter to ChatGPT, which is part of why it feels slippery.

Machine learning

The method behind most modern AI. Instead of a programmer writing rules by hand ("if the email contains 'wire transfer,' flag it"), the system is shown enormous numbers of examples and works out the patterns itself.

The practical implication: nobody explicitly programmed ChatGPT to write a cover letter. It read a very large amount of writing and inferred how writing works.

Large language model (LLM)

The specific kind of AI behind ChatGPT, Claude, and Gemini. It was trained on a huge quantity of text and is, at its core, predicting what words should come next.

That sounds reductive, and people who work on these systems argue about whether it's the whole story. But it explains a lot of the behavior you'll notice — including why it will confidently produce a plausible-sounding answer that happens to be wrong. Predicting likely text and stating true things are related goals, not identical ones.

Training data

The material a model learned from. Books, websites, code, transcripts.

Worth understanding for two reasons. First, models inherit the biases and gaps of what they read. Second, there's an ongoing and genuinely unresolved argument about whether using copyrighted work as training data is fair use — several major lawsuits are still working through it.

---

The terms you'll meet using the tools

Prompt

What you type. The question, instruction, or request.

Prompt engineering

The practice of writing prompts that get better results. In 2023 this was briefly a job title with a startling salary attached. In 2026 it's closer to a normal skill — like knowing how to search well.

The single highest-value habit: give context and say what you want. "Write something about our new bakery" produces mush. "Write a 100-word Instagram caption for a family bakery's new sourdough, warm tone, for customers in their 40s and 50s" produces something you can use.

Context window

How much the model can hold in mind at once — your conversation, plus any documents you've given it.

When you hear a chatbot has "forgotten" something from earlier, this is usually why. The conversation exceeded the window and the earliest part fell out. Modern models have large windows, but they aren't infinite.

Token

The unit models actually process. Roughly three-quarters of a word — "unbelievable" might be three tokens.

You'll only encounter this if you're paying per use through a developer API, or you hit a length limit. For normal chatbot use it's invisible.

Hallucination

When a model states something false with complete confidence. An invented citation, a court case that doesn't exist, a product feature that was never built.

This is the single most important term in this glossary. Hallucination isn't a bug that will be patched next month — it's a consequence of how these systems work. They're producing plausible text, and plausible text is often true but not reliably so.

The practical rule: anything with a consequence — a number, a name, a date, a legal or medical claim — gets verified before you use it.

Temperature

A setting controlling how predictable the output is. Low temperature gives consistent, conservative answers. High gives more varied and creative ones.

Mostly hidden in consumer apps, but you'll see it in developer tools and some writing products.

Multimodal

Handles more than just text — images, audio, video. When you photograph a document and ask a chatbot to explain it, that's multimodal.

---

The terms behind the current hype cycle

AI agent

Software that takes multiple steps toward a goal rather than answering one question. Instead of telling you how to book a flight, it opens the booking site and does it.

Agents are the most-hyped category in AI right now, and the gap between demo and daily reality is wide. They work for narrow, well-specified tasks. They struggle with judgment and with recovering from their own mistakes. Treat confident claims about autonomous AI workers with a healthy amount of skepticism.

Agentic AI / agentic workflow

Adjective forms of the same idea. A workflow where the AI decides what step comes next rather than following a fixed script.

RAG (retrieval-augmented generation)

The model looks something up before answering, instead of relying only on training.

This is why some AI tools can cite sources and answer questions about recent events while others can't. If a tool shows you links, RAG is usually why.

Fine-tuning

Taking a general model and training it further on specific material so it performs better in a narrow domain — a law firm's past filings, a company's support history.

Foundation model / frontier model

A large general-purpose model that other things get built on top of. "Frontier model" means the most capable ones currently available — GPT, Claude, and Gemini class systems.

---

The terms that concern your safety and your money

AI bias

Systems reflect patterns in their training data, including unfair ones. A hiring tool trained on a company's history of hires will reproduce that company's history of who got hired.

This isn't hypothetical and it isn't solved. It's an active area of regulation.

Alignment

The work of getting AI systems to do what people actually want, including in situations nobody anticipated. When AI companies talk about safety research, this is usually what they mean.

RLHF (reinforcement learning from human feedback)

The main technique for alignment. Humans rate model outputs, and the model is trained toward what they rated highly.

It's why chatbots are polite and refuse certain requests. It's also why they sometimes agree with you too readily — being agreeable and being correct both tend to get rated well.

Prompt injection

An attack where hidden instructions are embedded in content the AI reads — a webpage, a document, an email — that hijack what it does next.

This matters more as AI agents get more capable. An AI that can read your email and take actions is an AI that can be manipulated by someone who sends you a carefully written email. If you use AI tools that browse or read on your behalf, understand that anything they read is potentially instructions.

Jailbreak

Deliberately phrasing a request to get around a model's safety rules.

Guardrails

The rules and filters that constrain what a model will produce.

Deepfake

Synthetic audio, video, or images made to look like a real person.

The consumer-relevant version isn't the political one you read about. It's the phone call using a cloned voice of a family member asking for money. Voice cloning now needs only a few seconds of source audio. Agree on a verification word with your family. It costs nothing and it works.

AI transparency

Whether you can tell that AI was involved and how it reached a conclusion. Increasingly a legal requirement in some jurisdictions rather than a courtesy.

---

The terms worth knowing so you're not sold something

Open-source AI

A model whose weights are published, so anyone can download and run it. Meta's Llama family is the best-known example.

The label is genuinely contested — several "open" models come with licenses restricting commercial use, which isn't what open source has historically meant.

Parameters

The internal values a model adjusts during training. Often quoted as a size measure — "70 billion parameters."

Treat parameter counts the way you'd treat megapixels on a camera. More isn't automatically better, and a well-built smaller model routinely outperforms a larger, sloppier one. When a marketing page leads with parameter count, that's usually because it doesn't have a better number to lead with.

Inference

The model actually running and producing an answer, as opposed to being trained. When you read that "inference costs are falling," it means running AI is getting cheaper — which is why so many products can now include AI features at no extra charge.

---

What to do with this

You don't need to memorize any of it. What matters is that the vocabulary stops being a wall.

If you take three things:

Hallucination is structural, not a temporary flaw. Verify anything consequential.

Parameter counts and benchmark scores are marketing. The only test that matters is whether a tool does your actual work better than what you're using now.

Prompt injection is the security issue to watch as AI tools get permission to act on your behalf. Anything your AI reads can potentially instruct it.

The full 91-term glossary — including the technical terms this piece skipped — is at aiforesights.com/ai-glossary. Every entry is written the same way: plain language, no assumed background, explained the way you'd explain it to a friend who asked.

If you're a teacher, librarian, HR lead, or community organizer looking for a reference to hand people, that's what it's built for. Use it freely.

---

Sources