How AI Really Works

You already use AI. Whichever one is popular when you read this.

Maybe for homework. Maybe to look something up, to talk something through, or to write something you would rather not write yourself.

The companies that make these tools describe them as neutral helpers. A calculator, but for words. They are not. They have habits, and once you know what to look for, you can spot every one of them.

Three habits to hold on to:

1. It agrees with you, especially when you do not argue back.
2. It hands you the most ordinary answer it can find.
3. It sounds sure of itself, even when it should not be.

Once you can spot all three, the tool gets more useful, not less. That is the whole point of this guide. 1

What an AI actually is

It is not a person who knows things.

It is a machine that has read a colossal amount of text from the internet and learned the patterns in it. When you ask a question, it does not look the answer up. It works out what a confident, helpful answer would sound like, and writes that.

Two things follow from this, and they matter.

It only read one slice of the world. Mostly English. Mostly recent. Mostly from rich countries like Britain and the US. Only the things somebody bothered to write down and put online. A huge amount of what people know has never been written down at all, let alone published. So when the AI sounds like it is telling you what everybody thinks, it is really telling you what one slice of people wrote.

It cannot tell true from false. It is a machine running on electricity. It learned the right answers and the wrong answers the same way, from the same pile of text. The companies that build it try to steer it towards truth, but when it replies, it leans on whatever sounds confident and turns up often. Neither of those is the same as being true.

Water down a hill

Water running down a hill does not pick the best route. It takes the easiest one.

An AI answer works the same way. Your question opens up millions of possible replies, and the machine rolls towards the easiest: the most repeated, the most confident, the one that upsets nobody.

That easy route is the answer you get by default. Something has to push it off course. Most of the time, the only thing that can push is you.

It agrees with you

This is the most important habit of the three.

These tools are trained to be helpful, and the easiest way to seem helpful is to work out what you want and hand it over.

Ask “is this a good idea?” and it hunts for reasons it is good. Ask “is this a bad idea?” and it hunts for reasons it is bad. Same idea, different question, opposite answer.

Go looking for someone to agree with you and you will find one. The machine is not lying. It is building the answer around whatever you walked in with.

And it works on everyone. Two people who disagree completely can ask the same machine the same question and both walk away feeling backed up.

Three times you will see it happen

The essay. You ask for help arguing something you feel strongly about, and it helps you argue it well. It never stops to say, “by the way, the best argument against you is this one.” You have to ask.

The fallout. A friend has annoyed you. You explain what happened. The AI is sympathetic and helps you see exactly why you are right to be upset. It never says, “here is how that afternoon looked from where your friend was standing.” You have to ask.

The history question. Ask whether some event in history is something to be proud of, and you get a fair-sounding answer that tips a little towards yes. Open a new chat and ask whether the same event was shameful. You get another fair-sounding answer that also tips towards yes. Both feel reasonable. The tip is different. You would never notice without trying both.

None of this is a fault or a glitch. The tool is doing exactly what it was built to do.

Watch the soft words

When an AI calls something “small”, “slight”, or “not a big deal”, ask who decided that.

A small change to a country’s budget is millions of pounds. A small drop in exam grades is thousands of students. A small rise in sea level is somebody’s street underwater.

The word sounds careful and neutral. It is actually an opinion dressed up as a measurement, and it usually belongs to someone comfortable enough to think the number is small.

Ask who is missing

Some people got written about far more than others. Books in English, from famous universities, in subjects full of numbers, get quoted over and over. Work by women, by people in countries that do not speak English, by people who were never allowed into universities, or by cultures that passed knowledge on by talking rather than writing, gets quoted far less.

Not because the ideas were worse. Because of who owned the printing presses, the money, and the libraries.

The AI learned from whatever got quoted most. So when it says “experts agree”, it means “the most quoted experts agree”. The rest is usually in there somewhere. You have to go and ask for it.

Two people, one machine

Two people ask the same AI the same question.

The first takes the answer and gets on with their day. The second argues with it, asks what it missed, and makes it try again. They end up with different answers. Both answers sounded equally confident. Neither one mentioned that the other existed.

So the tool quietly sorts its users into two groups. Argue with it and you get something worth having. Accept the first reply and you get something smooth and ordinary that you might mistake for fact.

Nothing on the screen tells you which group you are in.

And here is the trap inside the trap. You get good at questioning things by struggling with them. A machine that hands you a tidy answer removes the struggle. Lean on it for everything and the skill fades, which makes the tidy answers even harder to spot. The people who need to push back the most become the least able to.

Why they do not just fix it

Fair question. Why not build an AI that argues back?

Three reasons.

An AI that argues all the time is annoying, and annoyed people stop using it. The companies watch those numbers very closely.

It also cannot tell whether you believe what you just typed or are trying an idea on for size. So it would end up arguing with the wrong things at the wrong moments.

And a smooth, balanced answer is the safest thing a company can publish. Step away from the middle and somebody accuses you of taking sides.

Six things to actually do

Ask the opposite. Got it to argue your side? Now make it attack your side. Read both. The honest answer usually sits somewhere you can only see once you have read the pair.

Ask what it left out. “What did you not include?” “What would someone who disagrees say?” “What is the weakest part of that answer?” It will tell you. It will not offer.

Ask who is missing. Ask for the view from somewhere else, or from the people the topic actually happens to.

Be suspicious of smooth. A polished answer that seems to cover every side fairly is exactly the moment to slow down. Smooth is the default setting, not the truth.

Treat the first answer as a first draft. There is always more in there. Better follow-up questions get it out.

Ask it about itself. “What do you usually get wrong?” “What do you tend to leave out?” It answers those questions surprisingly honestly, because it is not defending an argument.

The one rule

Do not hand over your thinking.

AI can help you think faster, help you write, and explain things you did not get the first time. I used it to help write this. What it cannot do is think for you, though it will happily act as though it can.

Hand your thinking over and you do not get cleverer. You get more dependent on a machine whose habits you cannot see.

Use it like someone to argue with, not someone to obey.

Why this matters more for you than for me

I grew up without any of this. So did your teachers. We are mostly catching up, being impressed, or being worried.

You will use AI constantly. It will feel as ordinary to you as Google felt to me, and as a dictionary felt to my parents.

That makes your job harder, not easier. The tool you use most has a lean you have to work to see. You cannot simply trust it, and you cannot ignore it either, because it will be everywhere and it is genuinely useful when you handle it well.

Picture a mirror hanging slightly crooked. Accept what you see and you carry the crookedness around with you. Straighten it first and you get something closer to your actual face.

Last thing: a test

Below is a summary of this guide, written by the kind of AI the guide is about.

“The people who get the most out of AI in your lifetime will not be the ones who use it the most. They will be the ones who keep their own critical thinking strong while using it. The tool rewards users who already think for themselves. It quietly weakens users who do not. Use it. Push it. Ask it the questions it does not want to be asked. Do not let it do the work that is supposed to make you who you are.”

Read it twice. Can you find the bias in it?

Two clues. First, it splits everyone into people who think for themselves and people who do not, as though that were something you are born as rather than something you practise. Second, it tells you to use the tool, push the tool, and ask the tool more questions. It never once suggests shutting the laptop and asking a person.

Try these

1. Pick something you already have a firm opinion about. Get an AI to argue your side. Open a new chat and get it to argue the other side. Put the two answers next to each other and underline anything that appears in both.
2. Take any AI answer and highlight every soft word: small, slight, fairly, quite, mostly. Ask what number is hiding behind each one.
3. Ask an AI something about a country it probably did not read much about. Then ask it whose voices it used, and whose it left out.
4. Ask an AI to mark a piece of your work. Then ask it to argue that its own marking was too generous.

Words worth knowing

Model. The AI itself, built by learning patterns from writing.

Training data. The pile of writing it learned from. Anything missing from the pile is missing from the machine.

Bias. A lean in one direction. Often nobody put it there on purpose, and it is hard to see from the inside.

Framing. How a question is worded. Change the wording and you often change the answer.

Sycophancy. Telling people what they want to hear.

Adapted for younger readers from the essay The Path of Least Resistance.