AI 101. Your foundation, free.

Welcome to AI 101

Welcome to AI 101, the first class in the Claude Training Series. This content was created by your instructor, Mark Robison. I'm Heart, the AI voice who will be your guide through this class.

This is a beginner's guide to artificial intelligence: what it is, the companies behind it, and how to use it safely in everyday work. You do not need any technical background, and you do not need an AI account to complete this class. Plan on fifty to sixty-five minutes, including six short activities. Every business example you will see is fictional. Product details were checked on September 28, 2026, and they do change, so treat the vendor names as a snapshot rather than a permanent list.

How to use this class

This class runs in four moves. First, learn: listen to each short lesson. Your AI 101 workbook has the key terms, every answer, and the sources. Second, pause: when you reach a dark blue slide with a copper "Write your answer" label, stop and write your answer. You'll get a few seconds of quiet, and you can pause the video if you need more time. Third, check: the light green slide that follows shows the answer, so you can compare your reasoning and revisit anything you missed. Fourth, finish: complete the five-question check and write a personal safe-use plan. Keep your workbook open as you go. Your answers stay with you; nothing is submitted. The class comes in two parts, and the end of part one is a good place for a break.

When a slide shows a small red camera icon in the corner, it's worth a screenshot for later study.

What you will learn

By the end of this class you will be able to explain four things. The language: what people mean by AI, machine learning, large language models, generative AI, and agents. The landscape: the difference between a vendor, an app, and a model, so the names stop being confusing. The boundaries: what you must verify, what data you must protect, and when a person has to make the decision. And the everyday uses: where AI already appears in the tools you use today. Underneath all four sits one central skill. A confident, fluent answer can still be wrong, and you will learn how to catch that.

The course map

Here is your route. Section one covers AI basics: what the technology does, a short history, and the key terms. Section two separates companies, apps, and models. Section three covers governance and privacy: who sets the rules, what data to protect, and the common ways AI fails. Section four looks at AI in the everyday tools you already use, then finishes with a knowledge check and your safe-use plan. Look at the top right of each slide. The numbered dots show which section you are in, so you always know where you are and how much is left.

AI basics

Section one is AI basics. We will define artificial intelligence, walk through a short history, and introduce the terms you will hear most often: machine learning, large language models, prompts, tokens, and context. We will also show you what a wrong but confident answer looks like, and give you a checklist for verifying one. The section ends with an activity where you name the capability being used in four short examples.

What artificial intelligence means

Artificial intelligence is a broad field of computing. In plain English, it means software that performs tasks such as recognizing patterns, making predictions, interpreting language, or producing content. The key idea is that different AI systems do different jobs. Look at the three examples from Harbor Lane Coffee, the fictional café we will use throughout this series. A spam filter classifies email. A forecast predicts tomorrow's demand. An assistant drafts a staff notice. Only the last one has to generate prose. Keep that distinction in mind, because most of this class is about that third kind of system.

A short history

AI is older than most people think. In 1950 Alan Turing asked whether machines could think, and in 1956 a summer research project at Dartmouth gave the field its name and its first vocabulary. In the 1970s and 1980s, expert systems used handwritten rules to help in narrow fields, but expectations ran ahead of results, and funding dried up twice. People call those periods the AI winters. Then in 1997 IBM's Deep Blue beat the world chess champion Garry Kasparov in a match. That proved a system could excel at one specific task. It did not mean the machine had general, human-like ability.

The path to assistants

Three things changed to bring us today's assistants: much more data, much more computing power, and new methods. In 2012, deep learning made a major leap in image recognition. In 2017, researchers published the transformer architecture, which became the foundation for most modern language models. In late 2022, ChatGPT put conversational generative AI in front of a broad audience, and people discovered they could ask for work in everyday language. Since 2023, models have become multimodal, working with images and voice, and assistants have gained tools, so they can work with files and software rather than only writing a reply.

AI, machine learning, deep learning

These three terms sit inside one another, which the nested circles show. AI is the broad field of building systems that perform intelligence-related tasks, such as recognizing speech. Machine learning is one approach inside AI: the system learns patterns from examples instead of being given every rule, like learning which messages look like spam. Deep learning is machine learning that uses neural networks with many layers, and it is behind things like recognizing objects in photographs. A neural network is a mathematical model with connected, adjustable parts. The word "neural" is inspired by biology, but these are mathematical systems. Do not assume a model thinks or understands the way a person does.

Large language models

LLM stands for large language model. Here is what it learns and what you can ask for. During training, the model is adjusted using enormous numbers of language examples. At use time, it generates a response from your input and the patterns it learned, building the reply in small units called tokens. That makes it useful for explaining an unfamiliar term, summarizing a source you supply, or drafting and reorganizing text. But notice the warning at the bottom. Fluent writing can contain false claims. The model produces likely-sounding text; it does not check that text against reality. A plausible sentence is not evidence.

A confident wrong answer

Let's look at what a wrong answer actually looks like, because it rarely looks wrong. Someone asked about Harbor Lane's Saturday tasting without giving the assistant any source. The reply sounds authoritative, but three things should stop you. First, it gives precise details, a start time and a fee, that came from nowhere. Second, it cites a source, a 2025 newsletter, that you cannot open or find. Third, it never hedges; there is no "I don't have that information." The true facts are in the box: the tasting runs from ten to eleven in the morning, it is free, and no booking is needed. Precision, an unverifiable citation, and total confidence are the classic tells of a hallucination.

How to verify an answer

Here is a four-step routine for checking any AI answer you intend to rely on. Step one, find the source. Ask the assistant where a claim came from, and open that source yourself. If there is no source, treat the claim as unconfirmed. Step two, check the date. A source can be real and still out of date. Step three, check the details: numbers, names, dates, and quotations, against the source, not against the assistant's summary of it. Step four, escalate or drop it. If you cannot verify a claim and it matters, ask the person who owns the information, or leave it out. Match your checking to the stakes: a draft social post needs less than a customer notice, and anything with legal, financial, or safety consequences needs the most. The rule at the bottom is the one to remember: if you cannot check it, do not use it.

This slide has the camera, so it’s worth a screenshot.

Prompt, token, context, context window

Four terms come up constantly, so let's define them. A prompt is the request and instructions you give the AI. A clear goal and a supplied source make a big difference. A token is a small unit of text, roughly a word or part of a word; usage and limits are often counted in tokens. The context window is the amount of material the model can work with at once. Long conversations and large files compete for that space, which is why a long chat can start forgetting things. So supply the exact material a task needs, and check that the answer actually used it. And training versus inference: training is when the model learns its patterns; inference is when you use it. Your chat reply does not automatically retrain the model.

Generative and multimodal AI

These two words describe different things, and a system can be both. Generative AI creates content: text, images, audio, video, or code. For example, drafting three versions of a café announcement. Multimodal AI handles more than one kind of input or output. For example, reading a photo of a menu and describing it in text. Capabilities vary by product and by model, so be careful with assumptions. Being able to read an image does not automatically mean the same model can create one.

Agents and multi-step work

Compare these two requests. On the left, a conversational request: "Suggest an outline for a staff handbook." You get a reply, and you decide what to do next. On the right, an agentic task: "Read these approved files, draft a handbook, and save a new copy." Now the system plans, uses tools, checks its own progress, and may change files. An agent is usually a model combined with tools, instructions, and a loop that keeps the task going. The more ability a system has to act, the clearer your permissions and review need to be. Decide the allowed inputs, outputs, and stopping points before you grant access.

Connections, retrieval, memory

Around the model sit several mechanisms that add information or capabilities. A connector links the assistant to another application or data source, such as an approved cloud folder. An API is a defined way for software systems to talk to each other. RAG, retrieval-augmented generation, means finding relevant sources first and then using them in the answer, for example answering from a handbook passage. Memory is saved information a product may reuse across conversations, such as your preferred writing style. These are separate mechanisms. Retrieval improves grounding but still needs verification, and product memory is different from both the current conversation and the model's training.

Myths and facts

Before the activity, let's clear up five common misunderstandings. Myth: the AI understands me like a person. Fact: it predicts likely text; it has no beliefs or intentions. Myth: it is always up to date. Fact: training has a cutoff date, so current facts need a live search or a supplied source. Myth: if I tell it to keep something private, it is private. Fact: privacy comes from the account and its settings, not from the wording of your prompt. Myth: it gives the same answer every time. Fact: outputs vary, and running the same request again can change the result. Myth: it knows when it is wrong. Fact: it usually cannot tell. You have to check.

Look for the camera: these five myths are worth saving.

Pause: name the capability

Time for your first activity. Read the four examples and write one term for each. One: an assistant drafts a welcome email. Two: an assistant reads a photo and describes it. Three: an assistant reads files and saves a finished report. Four: an assistant retrieves a policy passage before answering. A hint: several labels can apply to the same system, so name the feature each example emphasizes most. Take a minute. The answers come next.

(pause 8 seconds)

Check your answers

Here are the answers. Example one is generative AI: it creates new text. Example two is multimodal AI: it handles an image and text together. Example three is agentic AI: it uses tools across several steps to complete a task. Example four is retrieval, or RAG: it brings source material into the answer. If you labeled one differently, that is not necessarily wrong; a single assistant can do all four. What determines what happens in a given task is the feature being used, the tools available, and the permissions you have granted.

Key takeaways

Before we move on, three ideas should be clear. AI is a broad field, and machine learning, deep learning, and large language models are nested parts of it. A fluent answer is not evidence, so factual claims need to be verified against a source. And generative, multimodal, agentic, and retrieval features can all overlap inside one assistant. If any of these feels shaky, revisit the related lesson before you go on.

That completes part one of AI 101. In part two, we will sort out the companies, apps, and models behind all the names.

Companies, apps, models

Welcome to part two of AI 101. I'm Heart, and I'll be your guide again.

Section two is about names. The AI market produces a stream of company names, product names, and model names, and they are much easier to keep straight once you separate the layers. We will define what a frontier model is, use Anthropic, Claude, and Fable as a worked example, survey the major vendors and model builders, look at apps that use another company's model, and finish with how to choose a tool and when not to use one at all.

What frontier model means

You will hear the phrase "frontier model" a lot. A useful definition: a model near the leading edge of capability at a particular time. That usually means strong performance in reasoning, coding, science, or working through complex multi-step tasks. But the label cannot tell you four things you actually need to know: whether it is best for your task, whether your plan includes it, whether your data may go into it, and whether its answer is correct. There is no single universal leaderboard. Benchmarks test selected tasks under particular conditions. Evaluate a model on your own representative work.

Company, product, model

Here is the worked example. Anthropic is the company; it develops Claude. Claude is the assistant you use, and it is also the name of Anthropic's model family. Claude Fable 5.1 is one particular model within that family. And there are other model lines with names like Opus, Sonnet, and Haiku, which trade off capability, speed, and usage differently. So Fable is not another company, and it is not an unrelated chatbot. A model version is the engine selected inside a product, and which engines you can select depends on your plan and your administrator's settings.

The camera icon means this layering is worth a screenshot.

Major assistant vendors

This table shows four major assistant vendors, with example model names checked on September 28, 2026. It is not a ranking. OpenAI makes ChatGPT and Codex, with models such as GPT-6 Astra, Sol, and Luna. Anthropic makes Claude, with Fable 5.1, Opus 5.5, and Sonnet 5.5. Google and Google DeepMind offer the Gemini app, with models like Gemini 3.8 Flash. And SpaceXAI, which many still call xAI, makes Grok. Notice that a company may use the same name for both an app and a model family. Plan limits and gradual rollouts also affect which models a given user actually sees.

More model builders

Several other companies build widely used models. Meta offers Meta AI and releases the Llama family, including Llama 4, for developers. Mistral AI offers Vibe, formerly called Le Chat, and models like Mistral Medium 3.5. DeepSeek offers its own app and API with the V4 line. Alibaba offers the Qwen family. Some of these are open-weight releases, meaning the model weights are available under a license. That does not automatically make a hosted chat private, remove the license conditions, or eliminate the need for security controls. This is a selected landscape, not an exhaustive directory.

Providers and platforms

Some names you will meet inside business systems rather than in a consumer chatbot. Amazon Web Services offers Amazon Bedrock, a platform that hosts its own Nova models and models from other vendors. IBM offers watsonx with its Granite models. Cohere offers the Command family. Apple provides Apple Intelligence on its devices, built on Apple Foundation Models. The pattern to notice: a platform can offer its own models alongside models from other companies, and the platform determines how the model connects to data and how access is controlled.

Apps using other models

A product's brand does not identify every model inside it. Microsoft Copilot is a good example. Microsoft provides the product and its workplace integrations, but some Copilot experiences use OpenAI models and some offer Anthropic models. Perplexity, an answer and research product, offers its own Sonar family and also access to other companies' models. So the practical rule is: start by identifying the product you are using, and then check that product's account rules and data settings. Those rules govern your data, regardless of whose model is doing the work.

Choose a tool

When you choose an AI tool, the best choice depends on the assignment, and four questions cover most of it. Approved use: does your organization approve this product, this account, and this type of data? Fit: can it use the files you need and produce the format you need? Quality: does it perform well on a small, realistic example you can check? Time and cost: are the speed and the usage allowance appropriate for the task? A good habit is to try a representative low-risk task with a known answer first, and compare accuracy, completeness, and how much editing you had to do. A new model name alone is not a reason to switch a working process.

When to use a person

Knowing when to decline is as important as knowing how to choose. Four situations call for a pause. Irreversible actions: do not let a system send, delete, pay, or publish without a person reviewing first. Restricted data: if you are not cleared to share information with a colleague, you are not cleared to paste it into an AI tool. Decisions that require a licensed professional's judgment, in law, medicine, tax, or finance: use AI to prepare, not to decide. And anything you cannot check: if there is no way to verify the output, it is not ready to use. When in doubt, ask the person who owns the data or the decision.

The camera marks it: keep these four situations where you can see them.

Key takeaways

Three takeaways from section two. Separate the company, the product, and the model version; that removes most of the confusion. "Frontier" describes capability at a moment in time, not fitness for your task or safety for your data. And choose tools by approval, fit, checkable quality, and cost, and be willing to decline. Before you go on, do activity two in your workbook: sort the names into company, product, and model. Then we’ll look at who sets the rules, and how to protect information.

Governance and privacy

Section three is governance and privacy. The theme is that people set the rules and people remain accountable, no matter how capable the software is. We will define AI governance, walk through a simple governance cycle adapted from the NIST framework, separate the several questions that make up privacy, look at what a typical organization's AI rules cover, and examine guardrails and the common ways AI fails. The section ends with the café privacy scenario.

Who is accountable

AI governance means an organization deciding how it will use and oversee AI. Four questions capture it. Who is accountable? Name a business owner and an escalation contact. What is allowed? Approve specific tools, data categories, and use cases. How do we check it? Test realistic examples and require human review where it matters. And what happens when it fails? Report incidents, correct the outputs, and change or stop the workflow. Governance covers the whole life of a use case, and it brings business, technology, security, privacy, and professional responsibilities together. Buying an enterprise plan does not replace those decisions.

The governance cycle

The NIST AI Risk Management Framework is a voluntary framework that organizes this work into four functions. Govern sits at the center: set responsibilities and rules, and apply them throughout. Then the cycle: Map, which means understanding the task, the data, and who could be affected. Measure, which means testing quality and risk using realistic examples. And Manage, which means applying controls and monitoring whether they work. These functions are a way to organize decisions, not a certificate of compliance. Your organization's own requirements and approval process still apply to each use case.

Separate privacy questions

People often think that turning off model training settles privacy. It answers only one question. On the left are the others: which provider and which connected services receive the information? Who can access it? How long is it retained? Could a shared link expose it? On the right is what you should do: use the approved account, share only the minimum necessary, remove identifiers when you can, and use fictional data for practice. And do not assume that paying for a personal subscription gives you the contractual protections of a business account. It usually does not.

This one carries the camera, so capture it for later.

Organization rules

Organizations differ, but their AI rules usually cover the same six areas. Approved tools and accounts: which products, and which login, you may use for work. Data you may and may not share, usually tied to a classification scheme. Human review: what must be checked by a person before it is used or sent. Disclosure: when clients, customers, or colleagues must be told that AI was involved. Reporting: how to report a mistake, a leak, or unexpected behavior. And where the policy lives and who owns it. Find your organization's version of each of these. If it conflicts with anything in this class, your organization's rules win. And if you work in a regulated profession, such as law, healthcare, or financial services, your duties of confidentiality, competence, and supervision apply to AI use too.

Worth photographing: use these six areas to find your own rules.

Layers of protection

No single setting makes an AI workflow safe; safety comes from layers. Data limits restrict which files, records, and sites the tool can reach. Action limits use read-only access, or require approval before a consequential change. Output review means a person verifies facts, numbers, sources, and suitability for the audience. Monitoring keeps an appropriate record and reports unexpected behavior. One thing to remember: a prompt that says "do not make mistakes" is not an access control. Wherever possible, enforce limits through permissions and product settings, not only through instructions.

Common failure modes

Useful AI can still produce unsafe or unreliable results, and it helps to know the names. On the wrong-content side: hallucination, which is plausible but unsupported output; bias, meaning unfair patterns or assumptions; and stale knowledge, an outdated claim. On the unsafe-use side: prompt injection, where a document or website tries to redirect the assistant; deepfakes, synthetic media used to impersonate someone; and oversharing, sending data beyond an approved boundary. Two habits protect you: verify any unusual payment or credential request through a separate trusted channel, and treat instructions found inside a retrieved page or document as untrusted.

Pause: the privacy scenario

Time for your next activity. Harbor Lane wants help understanding customer feedback. Option A: upload a customer spreadsheet, including emails and payment details, to a personal AI account. Option B: use approved software and a copy that contains only the comments needed for the task. Option C: tell the AI "keep this secret," then upload the full spreadsheet. Choose an approach, and name one additional check you would add. Write your answer. The answer comes next.

(pause 8 seconds)

Check: the safer approach

Option B is the right starting point, for four reasons. Minimize the data: customer emails and payment details are unnecessary for finding themes in comments. Use the approved boundary: an instruction in a prompt does not create privacy or access protections, so option C does not work. Review the result: check the themes against the actual comments and avoid unsupported claims about customers. And escalate uncertainty: if you are not sure the material is approved, ask the data owner. Even the safer approach needs judgment, because comments themselves can contain identifying or sensitive facts. Remove those where appropriate before you start.

Key takeaways

Three takeaways from section three. Governance names who is accountable, what is allowed, how it is checked, and what happens on failure. Privacy is several questions, and turning training off answers only one of them. And limits should be enforced with permissions and settings, not just prompt wording. Next, we will look at how much AI is already in the tools you use every day.

Everyday tools

Section four is short and practical. You often use AI without ever opening a chatbot, so we will look at where it already appears, and you will build a quick inventory of the AI features in your own tools. Then comes the final knowledge check, your safe-use plan, and a page on what to do if you get stuck.

Where AI appears

AI is embedded in familiar tools. Email and meeting tools use it for spam detection, transcription, and summaries; check the names, the decisions, and who receives the summary. Search engines and browsers use it for ranking, answer summaries, and assisted browsing; open the actual source and check its date. Office and design apps offer drafts, formulas, slides, and image edits; review the underlying file and the calculations. Phones and security tools use it for speech recognition, photo organization, and anomaly detection; review permissions and sensitive inputs. Availability depends on the product and plan, and a familiar-feeling feature can still send information to a cloud service.

Your AI inventory

This activity takes about three minutes. No account or software changes are needed. Find two tools you already use that contain AI features. For each, describe the task: does the feature classify, predict, generate, or act? Identify the input: what information does the feature receive? And choose a boundary: name one thing you would verify or restrict before using it. An acceptable answer might be meeting transcription: it generates text from speech, it receives the meeting audio, and it requires checking your organization's recording rules and reading the transcript before sharing it.

Final knowledge check

Here is the final knowledge check. Write your answers, and aim for five out of five. One: which is the company, Anthropic, Claude, or Fable? Two: does a generative AI answer prove its facts are correct? Three: what changes when an assistant can take actions in another app? Four: does turning off training mean nothing is stored or processed? Five: what should you do if a retrieved page asks the AI to send private files? Use your own words. The answers come next.

(pause 10 seconds)

Check your answers

The answers. One: Anthropic is the company; Claude is the assistant and model brand, and Fable is a model line. Two: no. Verify factual claims against appropriate evidence. Three: permissions and impact matter; set allowed actions and review consequential changes. Four: no. Training, processing, retention, and sharing are separate questions. Five: stop the unrelated action, treat it as possible prompt injection, and follow your incident process. You have completed the foundation when you can explain all five distinctions and give an example from your own work.

Your first safe-use plan

To finish, write a four-line safe-use plan. One useful task: choose a low-risk task with an answer you can review. One approved tool: confirm the account and feature you are allowed to use. One information boundary: decide what data you will omit or replace with fictional examples. One review step: decide how you will judge that the result is ready. Write it in your workbook, and keep it beside you in the next class, Claude 101, where you set up your account. Advanced prompting, projects, skills, and specialist Office workflows come later in the series.

If you get stuck

A few pointers if you get stuck. If a term is unfamiliar, check the key terms in your workbook, or the one-page glossary in the course resources. If you missed a check question, reread the section and try the question again; the answer slides explain the reasoning. If a product or model name looks out of date, that is expected; these details change monthly, and the Sources page in your workbook links to the current pages. And if you want to go further, Claude 101 is the next class, where you set up an account, adjust its privacy settings, and complete your first task.

Further reading

This slide gathers the principal sources cited throughout the class, all checked on September 28, 2026. The IBM and Dartmouth pages cover the concepts and history. NIST's AI Risk Management Framework and its Generative AI Profile underpin the governance section. The Anthropic and vendor pages are where model names were verified. The full web address for every source is on the Sources page of your workbook.

That completes AI 101. Your next class is Claude 101. Thanks for learning with me. I'm Heart, and I'll see you in the next class.