When not to use AI for this

There is no shortage of writing about what these tools can do. There is very little about the cases where reaching for one makes the work slower, or worse, or both — which is strange, because everyone who uses them daily has hit all three.

What follows is not scepticism. I use these tools constantly and sell a few. It is the shorter, more useful list: the situations where I have learned to put the model down.

The first is when you cannot tell whether the output is right.

Every use of a model has two costs, not one. There is the cost of producing the thing, which is what gets advertised, and the cost of checking it, which does not. When you know the subject well, the second cost is small — you read the output, you see the two places it went wrong, you fix them, you are ahead. When you do not know the subject, the checking cost is the whole job. You are now reading something fluent and confident about a topic you cannot evaluate, which is a worse position than the blank page you started from, because the blank page was honest about what you did not know.

The tell is simple. Before you ask, try to describe what a wrong answer would look like. If you cannot, you are not ready to use a tool that sometimes gives one.

This is why "write me a contract" and "explain this contract" feel similar and are not remotely the same task. In the second, you are checking a summary against a document that sits in front of you. In the first, you are signing something you cannot audit.

The second is when the task is the thinking.

Some writing exists to transmit a conclusion you have already reached. Some writing is how you reach it. The two look identical on the page and are completely different activities, and only the first one can be delegated.

If you have ever written a long message to a colleague, worked out halfway through that you were wrong, and deleted the whole thing — that was the writing doing its job. A model will not do that for you. It will produce a fluent, well-structured version of the argument you asked it to make, including when the argument is bad. It has no stake in whether you are right and no way to notice that you are not.

Strategy documents, difficult emails, post-mortems, anything where you are still deciding what you believe: write these yourself, badly, and then use the tool to tighten what survives. The order matters more than the effort.

The third is when the input is the hard part.

A great deal of work that looks like generation is actually specification. The reason a page is hard to write is usually not the sentences. It is that nobody has decided who it is for, or what it is supposed to make happen, or which of the four things it currently claims is the real one.

A model will happily produce a page from an unclear brief. It will be well-formed and it will be about nothing in particular, and the vagueness that was in the brief will now be spread evenly across nine hundred words where it is harder to see. You have not solved the problem. You have made it longer.

You can tell you are here when you keep regenerating. Once is a bad roll. Three times means the instruction is the problem, and no amount of rephrasing the prompt will fix a decision you have not made. Stop, write down in one sentence what the thing is for, and most of the time you will find you can now write it yourself.

There is a fourth, and it is the one I get wrong most often: small tasks that are faster by hand.

Opening a chat, describing the context, waiting, reading, correcting, and pasting back is maybe ninety seconds of overhead. A surprising amount of work is under ninety seconds. Renaming a few variables, writing one email, checking one number. The tool is not slower at the task; it is slower at the whole loop, and the loop is what you actually pay.

The reason this one is sticky is that it does not feel like a loss. It feels productive. You were doing something the entire time.

What connects all four

In each case the model is doing the part that was never the bottleneck. Fluency was not the problem. Judgement was, or clarity was, or the overhead was.

That is the question worth asking before reaching for it, and it takes about five seconds: what is actually hard about this? If the answer is producing the words, or the code, or the first draft, these tools are extraordinary and you should use them without guilt. If the answer is knowing what should be produced, or being able to tell whether it is any good, you have not found a use for the tool. You have found a way to feel busy while the real work waits.

I would rather sell you something with that written on the box.

← All notes