What is AI?

Why This Matters

Every day, millions of people use AI tools without understanding what they actually are. Some treat AI like magic. Others dismiss it as hype. Both perspectives lead to bad outcomes -- you either trust it too much or miss real opportunities.

This module gives you the foundation. Not the marketing pitch, not the science fiction version -- the practical truth about what AI is, how it works, and what it means for you. Once you have the right mental model, every interaction with AI becomes more productive.

You don't need a computer science degree to understand this. You just need twenty minutes and a willingness to update your assumptions.

What AI Actually Is

Let's start with what AI is not. It is not sentient. It does not "want" things. It is not plotting anything. It is not a digital human trapped in a box. When you see AI described this way in movies, news headlines, or social media, you're looking at fiction or hype -- not reality.

At its core, modern AI is pattern recognition at massive scale. It finds statistical patterns in enormous amounts of data, then uses those patterns to generate new content that matches what it has learned. That is both less magical and more powerful than most people realize.

Types of AI

When researchers talk about AI, they usually distinguish between two categories:

• Narrow AI (what we have today) -- Systems designed to do specific tasks very well. Language models that write text. Image generators that create pictures. Code assistants that help you program. Each one is highly capable within its domain but cannot do anything outside of it. A language model cannot drive your car. An image generator cannot balance your budget.
• Artificial General Intelligence (AGI) -- A hypothetical system that could perform any intellectual task a human can. This does not exist yet. Serious researchers disagree about whether it will arrive in five years, fifty years, or ever. Do not let anyone tell you we are already there.

Every AI tool you can use today -- ChatGPT, Claude, Gemini, Midjourney, all of them -- is narrow AI. Remarkably capable narrow AI, but narrow nonetheless.

Multimodal AI

One important trend worth noting: modern AI is increasingly multimodal, meaning it can work with more than just text. Today's leading models can process images, audio, video, and documents in a single conversation. You can show an AI a photo of a whiteboard diagram and ask it to convert the sketch into structured notes. You can upload a chart and ask for analysis. You can describe an image you want and have one generated. This does not change the fundamental nature of AI -- it is still pattern recognition -- but it dramatically expands what these tools can do in practice.

How Large Language Models Work

The AI tools you are most likely to interact with -- ChatGPT, Claude, Gemini -- are all built on a technology called Large Language Models (LLMs). Understanding the basics of how they work will transform how you use them.

The Core Idea: Next-Token Prediction

Here is the fundamental concept: an LLM is trained to predict the next word (technically, the next "token") in a sequence. That's it. When you type a message to Claude or ChatGPT, the model reads your input and predicts what text should come next, one token at a time.

It sounds simple, but the results are extraordinary. To accurately predict the next word in a sentence, a model effectively needs to "understand" grammar, facts, reasoning patterns, tone, context, and more. Training on trillions of words forces the model to develop internal representations of how language and ideas work.

Tokens: The Building Blocks

LLMs do not read words the way you do. They break text into tokens -- chunks that are usually parts of words, whole words, or punctuation. The word "understanding" might be split into "under" and "standing." The word "AI" is typically one token. This tokenization allows the model to handle any text, including words it has never seen before, by working with familiar building blocks.

Training: Where the Knowledge Comes From

Before you ever type a message, an LLM goes through a massive training process. It reads enormous amounts of text -- books, websites, articles, code, conversations -- and learns the statistical relationships between tokens. This is called pre-training, and it typically requires thousands of specialized GPUs running for weeks or months.

After pre-training, models go through additional phases like fine-tuning (training on curated examples of good responses) and reinforcement learning from human feedback (RLHF), where human evaluators rate the model's outputs to help it learn what humans actually find helpful, accurate, and safe.

Temperature: The Creativity Dial

When a model predicts the next token, it doesn't just pick one answer -- it calculates probabilities for thousands of possible next tokens. A setting called temperature controls how the model chooses from these options:

• Low temperature -- The model almost always picks the most likely next token. Outputs are more predictable, consistent, and factual.
• High temperature -- The model is more willing to pick less likely tokens. Outputs become more creative, varied, and sometimes surprising.

This is why you can ask the same question twice and get different answers. The model is sampling from probabilities, not looking up a fixed answer.

Why AI "Hallucinates"

One of the most important things to understand about LLMs: they can generate text that sounds confident and authoritative but is completely wrong. This is called hallucination.

It happens because the model is always predicting the most plausible-sounding next token -- not checking facts against a database. If you ask about a topic where the training data is sparse or contradictory, the model will still generate fluent, confident text. It has no mechanism for saying "I actually don't know this well enough to answer." It just produces the most statistically likely response.

This is not a bug that will be "fixed" someday. It is a fundamental property of how prediction-based systems work. Every major LLM hallucinates. The frequency varies by model, topic, and technique, but it never reaches zero. This is why verification matters.

What AI Can and Can't Do

Having realistic expectations about AI is the difference between being productive with it and being frustrated by it. Here is an honest assessment as of early 2026.

What AI Does Well

• Drafting and editing text -- Emails, reports, articles, marketing copy, documentation. AI is an excellent first-draft machine and a tireless editor.
• Explaining concepts -- Translating complex ideas into simpler language, providing multiple explanations at different levels, creating analogies.
• Writing and debugging code -- Generating boilerplate, finding bugs, explaining error messages, suggesting optimizations, translating between programming languages.
• Analysis and summarization -- Condensing long documents, extracting key points, comparing options, identifying patterns in text.
• Brainstorming and ideation -- Generating options, exploring angles you hadn't considered, playing devil's advocate, stress-testing ideas.
• Translation and language tasks -- Converting between languages, adjusting tone, reformatting content for different audiences.

Where AI Struggles

• Factual accuracy -- AI can state incorrect things with total confidence. Always verify important facts independently.
• Mathematics -- While improving, LLMs still make arithmetic errors and logical mistakes, especially with multi-step problems.
• Real-time information -- Most AI models have a knowledge cutoff date. They do not automatically know what happened yesterday unless they have internet access tools.
• Persistent memory -- By default, AI does not remember previous conversations. Each new chat starts fresh unless the platform has built a memory feature.
• True reasoning -- AI can simulate reasoning patterns it has seen in training data, but it can stumble on novel logical problems, especially ones that require genuine multi-step deduction.
• Understanding your specific context -- AI does not know your company, your preferences, or your situation unless you tell it. It cannot read your mind.

The Right Mental Model

How you think about AI determines how effectively you use it. Here are three mental models that will serve you well.

Mental Model 1: The Brilliant Intern

Think of AI as a brilliant intern who has read everything on the internet. This intern has vast knowledge but no judgment. They can draft anything you ask for, recall obscure facts, and work at incredible speed -- but they will also confidently present wrong information, miss obvious context that you have not provided, and occasionally produce something subtly off-base. You would never publish an intern's first draft without review. Treat AI the same way.

Mental Model 2: A Tool, Not an Oracle

A hammer does not know what to build. A calculator does not know what to compute. AI does not know what you actually need. It is a tool -- the most flexible tool ever created, but still a tool. You provide the direction, the judgment, and the quality control. AI provides the speed, the breadth, and the first draft.

Mental Model 3: A Mirror of Your Input

The quality of AI output directly reflects the quality of your input. Vague questions get vague answers. Specific, well-structured requests get specific, useful responses. This is not a deficiency in AI -- it is exactly how a prediction engine should work. When you give it clear patterns to follow, it produces better predictions. When you give it ambiguity, it fills the gaps with whatever is most statistically likely, which may not be what you wanted.

Try This: Test Your Understanding