AI Foundations: How AI Works and the Words You’ll Hear

You can use ChatGPT, Gemini or Claude well without knowing how they work, the same way you drive without knowing how an engine works. But a teacher who understands the basics gets better results, trusts the tools in the right places, and can explain AI to students, families and colleagues who ask. This topic is that foundation: what today’s AI is, how it learns, why it makes the mistakes it makes, and the fifteen words you will keep hearing.

The AI You Already Use, and the AI This Course Is About

“AI” covers many systems. Your school already runs several: adaptive practice software that chooses the next problem, speech-to-text in a video call, spam filters, the grading engine behind a state test’s automated scoring, the recommendation that picks a video. Most of these are predictive systems: trained on past examples to sort, rank or score.

This course is about generative AI: systems that produce new text, images, audio and code on request. The models behind ChatGPT, Gemini and Claude are large language models, and the first thing to know is that recognizing which kind of AI you are dealing with is itself a core skill. Long and Magerko (2020) define AI literacy as “a set of competencies that enables individuals to critically evaluate AI technologies; communicate and collaborate effectively with AI; and use AI as a tool online, at home, and in the workplace,” and they list recognizing AI and telling narrow from general AI among the first competencies. Every tool in this course is narrow: excellent at language tasks, not a general mind.

How a Language Model Learns (in Four Sentences)

A model is built by training on an enormous collection of text (and now images, audio and video), far more than any person could read.
During training it learns one thing extremely well: given a stretch of text, predict what comes next, one small piece (a token) at a time.
The company then shapes its behavior with human feedback, so it answers questions, follows instructions, refuses some requests, and writes in a helpful tone.
When you type a prompt, the model generates a response token by token, choosing each next piece based on patterns it learned, plus everything currently in the conversation.

That is the whole mechanism. There is no database of facts being looked up, no understanding in the human sense, and no live memory of the world unless the product adds a search tool or you give it documents. Yiu, Kosoy and Gopnik (2024) argue that the best way to think about these systems is as a new cultural technology, like writing, print and the internet: “powerful imitation engines” that transmit and recombine what people have already written, rather than agents that seek the truth. Farrell, Gopnik, Shalizi and Evans (2025) make the same point for policy: large models generate summaries of “unmanageably large and complex bodies of human-generated information,” and they can reorganize that information in new ways, which is exactly why they are so useful for drafting and so risky for facts.

Why It Makes the Mistakes It Makes

Four everyday problems follow directly from the mechanism above:

Hallucination. Because the model predicts plausible text, it can produce a citation, a statistic or a quotation that reads perfectly and does not exist. Fluency is not evidence.
Training cutoff. The model learned from text up to a certain date. It knows nothing after that unless the product adds web search or you paste the new information in. (That is why this course’s model table carries a date.)
Bias. Patterns in the training text include the stereotypes and gaps of the people who wrote it. Generated examples, names and scenarios can carry them into your materials.
Confident tone. The model was shaped to be helpful, so it rarely signals uncertainty on its own. Good AI design asks systems to make clear what they can do and how well (Amershi and colleagues, 2019); until the tools do that reliably, the teacher supplies the judgment. That is the whole point of the Check the AI lesson.

Model Versus Product (and Why the Names Change)

Most confusion in workshops comes from one distinction. A model is the trained system with a version name: GPT-6, Gemini 4 Argon, Claude Opus 5.5. A product is the app you sign in to: ChatGPT, the Gemini app, Claude. The product wraps the model with memory, file uploads, web search, projects, skills and, increasingly, agents that act for you. Companies ship new models every few months and sometimes rename the products, so a feature name in this course may move; the ideas stay. When a plan says “includes GPT-6,” ask which version, on which plan, and in which part of the app. Several products now also let you choose a thinking level: more thinking means slower, more careful answers to hard questions, and a smaller quota.

The Words You Will Hear

Term	What it means for you
Model	The trained system (GPT-6, Gemini 4, Claude Opus 5.5). Changes every few months.
Product / app	ChatGPT, Gemini, Claude, Copilot: the model plus tools, memory and settings.
Prompt	What you type. The P-T-C-F recipe in the next lessons makes it work.
Token	The small piece of text the model reads and writes one at a time; roughly three-quarters of a word in English.
Context window	How much the model can hold in one conversation, including your files. Large today (about a million tokens on the leading models), but not unlimited.
Training data / cutoff	What the model learned from, and the date it stops.
Hallucination	A fluent, confident, false answer. Verify facts, quotes and sources.
Bias	Patterns from the training text that favor some groups, names or views.
Multimodal	Works with images, audio and video as well as text.
Reasoning / thinking level	A setting that lets the model work longer on a problem before answering.
Grounding	Answering from documents you supply (Gemini Notebook, Projects) instead of memory. Fewer invented facts, still worth checking.
Agent	AI that takes multi-step actions toward a goal: researching, writing files, using apps (Lesson 5).
Skill	Reusable instructions the AI applies on its own when a task matches. The word all three vendors now use (Lesson 4).
Plugin / connector	A link between the AI and an outside service such as Drive, Gmail or Canva, often built on a shared standard (MCP).
Open vs. closed model	Whether the model’s weights are published for anyone to run (open) or available only through the company’s product (closed, like the three in this course).

A One-Minute History

Late 2022: ChatGPT makes a language model available to everyone through a chat box. 2024: the leading models become multimodal and their context windows grow to hold whole books. 2025: products add memory, projects, deep research and the first agents that browse and act. 2026: the three vendors converge on “skills” as the reusable-instruction standard, fold agents into the main products (ChatGPT Work, Claude’s unified assistant, Gemini’s Spark), and ship free programs for verified teachers. This course is updated each quarter because that pace has not slowed.

AI Literacy as a Professional Competency

UNESCO’s AI Competency Framework for Teachers (2024) describes 15 competencies in five dimensions: a human-centred mindset, ethics of AI, AI foundations and applications, AI pedagogy, and AI for professional learning, each at three levels (acquire, deepen, create). In our reading of that framework, this course works mainly in AI foundations and applications and AI for professional learning, at the acquire and deepen levels, and the Using AI Responsibly lesson addresses the ethics dimension. If your district or university is building an AI plan, that framework is the reference point most countries now start from.

Quick Practice: Explain It in Two Sentences

Write, in your own words, how a language model produces an answer and why it can be wrong. Then paste your two sentences into ChatGPT, Gemini or Claude and ask: “Is this accurate? What would you correct?” Notice that you are now able to judge its correction.

Wrap-Up

Today’s AI tools are narrow systems that predict text from patterns in what people have written, shaped by human feedback into helpful assistants. That one idea explains their strengths (drafting, summarizing, translating, reorganizing) and their failures (invented facts, stale knowledge, inherited bias, confident tone). Keep the model-versus-product distinction in mind, learn the fifteen words, and you can follow every announcement that comes after this course.

Sources: Long, D., & Magerko, B. (2020). What is AI literacy? Competencies and design considerations. CHI 2020. · Yiu, E., Kosoy, E., & Gopnik, A. (2024). Transmission versus truth, imitation versus innovation. Perspectives on Psychological Science, 19(5), 874–883. · Farrell, H., Gopnik, A., Shalizi, C., & Evans, J. (2025). Large AI models are cultural and social technologies. Science, 387(6739), 1153–1156. · Amershi, S., et al. (2019). Guidelines for human-AI interaction. CHI 2019. · UNESCO (2024). AI competency framework for teachers.

Updated: October 2026

Previous Module

Back to Module

Next Lesson