0.2 · AI in 2 minutes, no jargon

In 2 minutes

An artificial intelligence is a program able to carry out tasks we thought were reserved for humans: writing, summarising, answering questions, translating, creating an image.

The truest image is that of the very fast, very well-read intern. It has read an enormous amount, it works in seconds, it never tires. But it needs clear instructions, it does not know your situation unless you tell it, and it can be wrong with disconcerting confidence.

The important point, the one that changes everything: an AI does not consult a store of truth. It works out what is most likely, word after word, from everything it has read. Most of the time, the most likely happens to be true. Sometimes not.

The image of the very fast intern still holds. But to know when to trust it, you need four more things. They will not change when the models change.

1 · It writes, it does not look things up

You read it above: it works out the most likely sequence, word after word. Of its own accord, it does not go and look for the answer anywhere. And even when it has first run a web search, it is still the one writing the answer, word after word, from what it found. Now add the consequence, which is the real lesson of this chapter.

The same mechanism produces its fine sentences and its inventions. It is not a flaw alongside a strength, it is one single thing seen from two sides. You will never be able to keep the one and remove the other.

2 · It has read a great deal, but very unevenly

It knows what it read during its training, and that training stopped at a date. Without a tool, it knows nothing of what happened afterwards: what you give it, or what it goes and finds on the web, fills that gap.

But the right question is not “does it know this subject?”. It is: has the world written a lot about it? On writing a business letter, millions of pages exist: it is excellent. On the tax regime for micro-businesses in Benin, it has read very little: it stays fluent, and it becomes fragile. The tone does not change. The reliability does.

The same goes for languages: it sometimes answers better in some than in others. What has been little written has been little read.

3 · It only sees one window

It pays attention to what you have put in front of it: the current conversation, the attached files. That space is large. It is finite.

This limit is not like the other three. The other three degrade gently, and you feel it coming. This one gives no warning: it works, it works, and then it stops working, often without any message telling you so.

And one detail that explains half the disappointments: it reads the beginning and the end better than the middle. You paste in a thirty-page commercial lease, and the clause you care about is on page seventeen: there is a good chance it will miss it. Put the essentials at the top, and say them again at the bottom.

Hence a rule that runs against intuition: giving it more does not always improve the result. Every document you add pushes the others towards the middle, that is, towards the place where it looks least closely. Two well-chosen pages are worth more than forty.

And what the window does not keep: from one conversation to the next, the model itself remembers nothing, and your corrections do not change it. They hold for the current conversation. What carries over from one conversation to the next is what the application puts back in front of it: its memory, your projects, your instructions. You will set them up in chapter 1.3.

4 · It follows the words, not the intention

It obeys short, checkable instructions remarkably well: “in three bullet points”, “under a hundred words”, “as a table”. You are far more in control than people think.

But it obeys the way it does everything else: by extending what you wrote. So there will always be a gap between what you meant and what you said. When it obeys you to the letter and the result is useless, do not repeat the instruction louder: tell it what it is for. Not “make it shorter”, but “the client has to understand my request from the very first line”.

And there is one thing it does badly by nature: exact calculation. For a total, an average, a percentage, have it done by a tool (a spreadsheet, or a calculation it runs itself) or do it again yourself. The same caution applies to reasoning in several steps: it can drift along the way. Ask to see step 2 before it carries on.

Let's do it together

Nothing to install for this chapter. Three ideas to anchor, and they will serve you through all the rest of the handbook.

1. Read this sentence twice: an AI predicts a plausible answer, it does not consult a store of truth.
2. Remember the consequence. When it does not know, it does not stop: it invents something credible. This is called a hallucination. It is not a rare bug, it is normal behaviour pushed to its limits.
3. Remember the reflex. For anything that really counts, a figure, a date, a reference, a point of law, a piece of medical information: you check elsewhere. For the rest, you can move fast.

The seven words you will come across

Before going further, seven words. Seven, not fifty: these are the ones that turn up everywhere from the first hour, and knowing them will stop you getting stuck on a menu or a message.

What you do not hand over to an AI

You will soon be putting your own documents into Claude, and that is the whole point. One simple rule before you go there: what you write is sent to the publisher's servers. There is nothing worrying about that for a set of meeting notes. There is rather more for these five things.

• Passwords, codes, access keys. Never, whatever the tool.
• Identity documents and official numbers.
• Health data, yours or anyone else's.
• Personal data belonging to your clients or colleagues that you have no right to pass on. Anonymise before pasting.
• Professional secrets covered by a confidentiality agreement.

For everything else, your notes, your drafts, your ordinary working documents: go ahead. And do go and look once, in your Claude account settings, at the section on using your data to improve the models: the setting exists, it changes with one click, and it is better to have seen it now than to discover it in six months.