An AI glossary is just a list of definitions for the words that get thrown around whenever people talk about artificial intelligence. That sounds simple, and yet most glossaries are weirdly hard to use. They sort terms alphabetically, so "deep learning" sits four screens away from "machine learning" even though you cannot really understand one without the other. You look up a word, get a definition stuffed with three more words you do not know, and leave more confused than you arrived.
This piece is built differently. Instead of an A to Z dump, the terms are grouped the way they actually relate to each other, so by the time you reach the bottom you have a mental map and not just a pile of cards. The goal is modest and practical: when a vendor demo, a district memo, or your own teenager uses one of these words, you should know what they mean and roughly whether to be excited, skeptical, or unbothered.
Start with the nesting dolls
Four terms get used as if they were synonyms. They are not. They nest inside each other like Russian dolls, and getting the nesting right clears up most confusion before it starts.
Artificial intelligence (AI) is the broadest term. It is the umbrella for any technique that lets a computer do something we would call "intelligent" if a person did it: recognizing a face, recommending a video, answering a question. AI is the whole field, not a specific product.
Machine learning (ML) is a subset of AI, and it is where almost all the action is today. Instead of a programmer writing explicit rules ("if the email says 'free money,' mark it spam"), the system is shown thousands of examples and figures out the patterns on its own. The "learning" is just statistics at scale. The machine is not understanding anything. It is finding correlations in data and using them to make predictions.
Deep learning is a subset of machine learning that uses neural networks with many layers, which is where the "deep" comes from. The extra layers let the system handle messy, high-dimensional things like images, audio, and language. Deep learning is the reason AI got suddenly good in the 2010s. It is powerful and also a bit of a black box, which we will come back to.
Generative AI is the newest layer most people are reacting to. It is deep learning aimed at producing new content: paragraphs, essays, images, code, songs. ChatGPT, Claude, and Gemini are generative AI. When a student turns in an essay that feels off, generative AI is almost always what produced it. The key word is generative. Older AI mostly sorted, scored, or predicted. This kind creates.
Hold onto the nesting: generative AI is a kind of deep learning, which is a kind of machine learning, which is a kind of AI. Every one of those phrases describes the same family at a different zoom level.
How the language models actually work
The tools dominating education conversations are large language models, or LLMs. It is worth understanding the mechanism, because it explains both why they are impressive and why they fail in such specific ways.
A large language model is trained on a staggering amount of text and learns, at its core, one deceptively simple skill: predicting the next chunk of text given everything before it. Ask it a question and it generates a plausible continuation, one piece at a time. That is it. There is no database of facts it looks things up in, no understanding in the human sense. It is an extraordinarily sophisticated autocomplete.
A token is the unit it works in. Tokens are pieces of words, roughly four characters of English each. "Unbelievable" might be three tokens. Models read and write in tokens, and limits are measured in them, which is why a tool might cut off mid-thought when a document gets long.
A prompt is whatever you type to the model. Prompt engineering is the craft of phrasing that input to get a better result. It is less a science than a knack, and it is the single most useful AI skill a teacher or student can build, because the same model gives wildly different answers depending on how you ask.
A parameter is one of the internal numbers the model tunes during training. Big models have hundreds of billions of them. You will see parameter counts cited like horsepower, but more parameters do not automatically mean better answers for your purpose, so treat the number as trivia rather than a buying signal.
The most important term in this whole section is hallucination. When a model states something false with total confidence, that is a hallucination. It is not lying, because lying requires knowing the truth. The model is just generating a plausible-sounding continuation that happens to be wrong. Fake citations, invented quotes, confident wrong dates: all hallucinations. This is not a bug that will be fully patched away. It is baked into how the technology works, and it is exactly why "the AI said so" can never be the end of a conversation in a classroom.
The terms that show up in vendor pitches
Walk through an ed-tech exhibit hall and you will be bathed in a particular set of words. Most of them are real ideas wrapped in marketing gloss. Here is the translation.
AI engine sounds technical but usually just means "the AI part of our product." There is no standard definition. When a vendor says their platform has a proprietary AI engine, ask what model it is built on and what data it was trained on. The answer tells you far more than the brand name.
AI-enabled personalization is the promise that software will adapt to each student, serving easier or harder material based on performance. The idea is genuinely good. The execution varies enormously, and "personalization" sometimes means little more than a difficulty slider with a fresh coat of paint. Worth asking for evidence, not just the claim.
AI analytics refers to using these techniques to find patterns in student data: who is falling behind, which concepts a class is struggling with, when someone is at risk of dropping off. Useful when the underlying data is good. Dangerous when a dashboard's confident-looking number gets treated as truth rather than a flag worth investigating.
Data mining is the older, broader practice of digging through large datasets to surface patterns and relationships. It predates the current AI wave by decades. In a school context it raises the obvious question every parent should ask: whose data, collected how, stored where, and shared with whom.
Training data is the material a model learned from, and it deserves its own line because it explains almost every problem downstream. A model trained mostly on formal published English will struggle with, and sometimes penalize, writing that does not sound like that. Bias in, bias out. When you hear that a detector or a model is unfair to some group of students, training data is usually the root cause.
The words that matter most for fairness
Two clusters of vocabulary carry real weight for anyone responsible for students, and they are the ones marketing tends to skip.
Bias, in AI, does not mean a grudge. It means a systematic skew that makes the system reliably wrong in one direction, usually inherited from imbalanced training data. A grading model that scores certain dialects lower is biased, even though no person sat down and decided to do that. The harm is real regardless of intent, which is why "the algorithm did it" is never an excuse.
Black box describes a system whose internal reasoning you cannot inspect. Deep learning models are notoriously black-box: they produce an answer, but they cannot show their work in any way a human can audit. This matters enormously when a tool flags a student for cheating or assigns a grade. If the system cannot explain itself, the human using it has to carry the full weight of the judgment.
Explainability is the push back against the black box: building or choosing tools that can say why they reached a conclusion. In education this is not a luxury. A plagiarism or AI-writing flag that comes with reasons a teacher can evaluate is categorically more trustworthy than a lone percentage with no context behind it.
AI ethics in education ties the cluster together. It is the practice of asking who benefits, who could be harmed, whose data is involved, and what happens when the tool is wrong, before deploying anything at scale. It is not a compliance checkbox. It is the difference between a school that uses these tools thoughtfully and one that lets a vendor's defaults make consequential decisions about children.
A few misconceptions worth retiring
"AI understands what it writes." It does not. It predicts text. Fluent output is not evidence of comprehension, which is why a model can write a gorgeous paragraph and get the underlying facts completely wrong.
"More data and more parameters always mean a better tool." Bigger is sometimes better and sometimes just bigger. For a school, fit, transparency, and evidence matter far more than raw scale.
"AI is objective because it is math." Math built on skewed data produces skewed results with a veneer of neutrality. The numbers feel impartial, which is precisely what makes unexamined bias so dangerous.
"Detection tools are simply the opposite of generation." They are related but not mirror images. A detector estimates the probability that text was machine-generated. It produces a likelihood, not a verdict, and like any model it can be wrong, which is why its output belongs in a conversation with a student rather than at the end of one.
What to do with all this
You do not need to memorize a glossary to be the smartest person in the room about AI in your school. You need the map. AI contains machine learning contains deep learning contains the generative tools your students are using. Those tools predict text, which makes them fluent and unreliable in equal measure. Everything they do traces back to training data, which is where both their power and their bias come from. And the words vendors love most, engine and personalization and analytics, are worth exactly as much as the evidence behind them.
Vocabulary is not the point. Good questions are. But you cannot ask a good question about a word you do not understand, and now you understand the words. Next time someone drops "the AI engine uses deep learning for personalization," you will hear it for what it is: a sentence that means something, and one you are fully equipped to push on.