Chapter 1: AI Is Already Here
⚡ FAST TRACK
You have been using AI for years
You are closer to using AI well than you may think. Long before your first prompt, intelligent systems were already helping with familiar parts of your day.
It is not new, and you have been using it for a long time.
When your email inbox catches spam before you see it, that is a machine learning system trained on hundreds of millions of messages. When your phone corrects “teh” to “the,” when it suggests the next word, when it finds every photo of your dog by searching the word dog, those are trained models. When your bank texts you about a suspicious charge in a city you have never visited, a fraud-detection model flagged it. When your maps app reroutes you around traffic that has not yet appeared on the road in front of you, that is prediction from data. When a streaming service recommends something you actually like, that is a recommendation model that has studied your behavior and the behavior of millions of people similar to you.
None of that felt like a decision. It arrived quietly inside products you were already using, and it made those products better without asking you to learn anything.
So when people say they are “not sure they want to get into AI,” what they usually mean is something narrower and more specific: they are not sure about the chat kind. The kind you talk to.
Fair enough. That kind is different, and it is worth understanding why.
What changed
The AI embedded in your spam filter is narrow. It does one thing. It was built for that one thing, trained on data for that one thing, and it cannot do anything else. A spam filter cannot summarize a contract. A fraud model cannot write a condolence note.
What changed, starting around 2020 and accelerating hard from late 2022, is that we learned how to build systems that are general. One system, trained on an enormous quantity of human writing, that can attempt almost any task expressible in language. Summarize this. Explain that. Draft this. Translate that. Find the flaw in this reasoning. Turn these notes into a plan.
These are called large language models, and they are the engine powering ChatGPT, Claude, Gemini, Copilot, and most of today’s major chat-based AI tools. For now, one sentence will do: they learned the patterns of human language by processing a staggering amount of text, and it turns out that predicting language well requires learning a great deal about the world the language describes.
The second thing that changed is the interface. For decades, using a computer meant learning the computer’s language: menus, syntax, file paths, formulas, commands. Now, for the first time, the computer meets you in yours. You type what you want in ordinary English and something useful comes back.
The actual revolution came when the cost of asking a machine for help dropped to almost nothing.
I studied artificial intelligence in graduate school long before any of this was in the news, back when the field’s ambitions and the field’s results were separated by a wide and slightly embarrassing gap. I want to be careful not to overclaim here, because overclaiming is the standard failure mode of everyone writing about this subject. But I will say this: the gap closed faster than most of us expected, and it closed in a direction nobody in my program predicted.
Three honest ways to think about what you are dealing with
Metaphors for AI are all wrong in some way. These three are wrong in useful ways.
A very well-read assistant who has never worked here. It has read more than any person could and can discuss nearly anything. It also does not know your business, your history, your constraints, or your standards, and it will confidently guess rather than admit it is guessing. Everything you would tell a capable new hire on their first day, you have to tell this thing every time. This metaphor is the one to hold when you are deciding how much context to include in a request, which is nearly always more than you think.
A calculator for words. A calculator does arithmetic faster and more reliably than you do, and it does not relieve you of understanding the problem. Punch in the wrong numbers and you get a wrong answer instantly and confidently. Language models are the same instrument for a different kind of work: enormously fast, indifferent to whether the input made sense. This metaphor is the one to hold when you catch yourself thinking the output must be right because it came out so smoothly.
A confident stranger. Imagine asking a well-spoken stranger at a party a question in their supposed field. Most of what they say will be broadly right. Some of it will be subtly wrong. All of it will sound equally certain, because the fluency of the delivery has nothing to do with the accuracy of the content. You would verify anything important before acting on it. Do that here. This metaphor is the one to hold at the moment you are about to send, submit, or sign something.
Together, these three metaphors suggest a productive relationship: give AI context, use its speed, and verify important results. That combination is the foundation for confident use.
Why “wait until it settles down” doesn’t work
Rapid change can make waiting seem sensible: let the dust settle, then learn whichever version survives. The more useful approach is to learn the durable skills now.
I understand the logic and I think it fails, for three reasons.
It will not settle. Software of this kind does not reach a final form and stop. Word processors have been changing continuously for forty years. The people who waited for word processing to settle down are still waiting, and they spent those decades at a disadvantage.
The durable skills are available right now. How to frame a request, how to check an answer, how to decide what is safe to share: none of that depends on which version of which product is current. You can learn all of it today, and it will still be correct when the products look completely different. That is what this book is.
The gap compounds. People who started using these tools two years ago are not two years ahead on features. They are two years ahead on judgment: they know what to delegate, what to check, what never to hand over, and how to notice when an answer smells wrong. That kind of knowledge only accumulates through use, and you cannot compress it by reading faster later.
Starting now gives you time to develop judgment through experience.
Where this technology shines
A few capabilities deliver useful results almost immediately. Knowing them will help you begin with work that feels rewarding.
Language models are strong at transformation: taking something that exists and changing its form. Long to short. Technical to plain. Notes to prose. English to Spanish. Rough to polished. This is where they are most reliably excellent, and where verification is easiest, because you have the original in front of you.
They are strong at generating options. Twenty subject lines, ten names, five approaches to a difficult conversation. Quality varies, but volume plus your judgment beats staring at a blank page.
They are strong at explanation. Ask for something to be explained at any level, in any style, with any analogy, as many times as you need, with no impatience and no judgment. Patient, adaptable explanations have quietly changed how many people learn.
They are strong at structure. Turning a mess into an outline, a plan, a table, a checklist.
These strengths become even more useful when paired with one simple habit: check consequential facts. Numbers, dates, citations, quotations, legal provisions, and medical dosages deserve confirmation at a reliable source. Verification is part of skilled use, much as checking a calculation is part of skilled engineering.
They have no access to your world unless you give it to them. Your files, your calendar, your company’s data, this morning’s news: none of that is inside the model. Some tools can now reach out to fetch such things, which helps enormously.
And they have no judgment. They will help you write a resignation letter without asking whether resigning is wise.
What you should expect from yourself in the first week
Expect a capable assistant rather than a perfect authority, and give your own judgment an active role. That combination works remarkably well.
Your first attempts may produce ordinary output, and that is completely normal. Most beginners start with requests that are too short to reveal what they need. Add context, try again, and you will often see an immediate improvement. Later, you will turn that habit into a repeatable method.
Expect to iterate. Nobody good at this gets what they want in one exchange. They get something close and then say shorter, less formal, drop the third point, give me three more like the second one. The conversation is the tool, not the first reply.
Expect to verify anything you will act on, send, submit, or attach your name to. Routine transformations of your own text rarely need fact-checking, but consequential claims do. That habit determines whether AI makes you more effective or eventually makes you look foolish in public.
Exercise 1.1: Find the AI already in your day
Before you open any AI tool, spend ten minutes on an inventory.
List every place you can identify where software already makes a prediction or judgment on your behalf: your email, phone keyboard, camera, bank, maps, streaming services, online shopping, workplace software.
For each one, answer two questions:
1. What would I lose if it stopped working tomorrow?
2. Have I ever checked whether it was right?
Most people find they rely on a dozen automated judgments daily and verify none of them. That is usually fine; the stakes are low and the systems are narrow. The purpose of this exercise is to notice the habit, because you are about to start using a far more capable system on far higher-stakes work, and the same unexamined trust will not serve you there.
Exercise 1.2: Write down your skeptical question
Everyone approaching this technology has one real objection. Not the polite one — the real one. Will it take my job? Is it stealing from writers? Is it making people stupid? Is any of this actually true or is it all a bubble?
Write yours down on the inside cover of this book. I have tried to address the serious versions of all of those, and I will point out where I do. At the end of the book, come back and see whether the answer satisfied you. If it did not, that is worth knowing too, and it means you were reading the way I hoped you would.