The short version
An AI assistant is a computer program that has read an enormous amount of writing. Books, articles, websites, transcripts, arguments, recipes. From all of it, it has learned in extraordinary detail which words tend to follow which other words.
That is genuinely all it is. There is no small person inside. It is not thinking about you between your questions. When you type something, it works out, a word at a time, what a good answer would look like.
That sounds like it ought to be useless. It is not, and this is the part that surprises people: once a program has read enough, “what a good answer would look like” turns out, very often, to be a good answer.
Everything else on this page follows from that one fact. What it does well, it does well because of it. Where it fails, it fails because of it.
What it is genuinely good at
- Explaining things at whatever level you want. You can ask for the simple version, then the slightly less simple version, then an example. It will not get impatient and it will not think less of you.
- Putting words together. A letter to the city. A note to your daughter’s landlord. A polite complaint. Tidying up something you have already written.
- Shortening things. Long article, dense letter, forty-page document. Ask what it says and it will tell you.
- Being a first draft of a thought. Ideas for a talk, questions to ask a doctor, ten things to do in a city you are visiting.
- Repeating itself without sighing. Ask the same question five times, in five slightly different ways, until it lands. This is the underrated one. Nobody else in your life will do this for you.
What it cannot do
It does not know anything about you that you have not told it. Not your bank balance, not your appointments, not your address.
In an ordinary chat it cannot do anything out in the world. It cannot send an email, move money, or book a flight. It writes text. That is the extent of it.
It does not necessarily know what happened last week. Some assistants can look things up on the internet and will tell you when they have. Many answers come from what the program read during its training, which stopped at some point in the past.
And it can be wrong. Which brings us to the thing worth remembering.
Why it sometimes sounds more certain than it should
An AI assistant has no inner voice saying I am not sure about this one.
When it knows something well, it produces smooth, confident, well-organized text. When it does not know something, it produces smooth, confident, well-organized text. The same machinery makes both. The tone is not a signal. It cannot be, because nothing about how certain it sounds is connected to whether it is right.
This is not the machine lying to you. Lying would require knowing the truth. It is closer to a very well-read person who cannot bear to say “no idea,” and who fills the gap with something that has the right shape.
You will hear people call this a hallucination. That is the word for it: an answer delivered in the same steady voice as all the others, that simply is not so. Invented book titles. A date that is two years off. A rule that does not exist. Confidently, pleasantly wrong.
Knowing this is most of what you need. It changes nothing about how you use the thing day to day. It changes what you do with the answer.
The rule worth carrying
Trust it most where you could check it. Check it wherever it would matter.
Ask it to explain how a heat pump works and you will get a good explanation. Ask it to rewrite your letter and you can read the letter and see for yourself. Nothing needs verifying, because you are the judge.
But check the specific things against the actual source: a dose, a legal deadline, a bank’s phone number, a claim about a medicine, a figure you are about to repeat to someone else. Not because it is usually wrong about them. Because you cannot tell from the answer whether this is one of the times.
That is the whole of it. Not fear, not suspicion. Just knowing which half you are in.
Try this now
Pick something you know extremely well. Thirty years of your working life. The town you grew up in. Your hobby. The thing your family asks you about.
Ask the assistant to explain it to you.
Then read the answer as the expert you are.
Most people find two things at once. It gets a surprising amount right, more than they expected. And somewhere in there is a detail that is off. Small, plausible, said in exactly the same voice as everything else.
That is the whole lesson, and you cannot really learn it by being told. You have to catch it once yourself.
If you find the error, say so. Type that is not right, [the thing] is actually [the thing] and see what it does.
If something did not go as expected
- It got everything right and I found no errors.
- That happens, particularly on well-documented subjects. Try something more specific: a local detail, a small organization, an exact date. The gaps live in the corners.
- It gave me far more than I wanted.
- Type shorter, please or just the main points. It will adjust immediately.
- I have not opened an assistant yet.
- Then start with Getting Started, which takes about fifteen minutes, and come back to this page afterward. Nothing here depends on doing them in order.
- I corrected it and it just agreed with me, even though I was testing it.
- Yes. It does that, and it is worth knowing. It tends toward agreeableness. Being firm with it is a real skill, and it is covered in Step 4.