I wrote about how Moravec’s Paradox defines what AI is good at, and what it’s not good at. Essentially, it can do things that we can reduce to symbols (calculating, chess, build LLMs) but not those things we humans find hard to put into symbols (passing a delicate item to another person, choosing what to draw).

With pressure to get AI to do my job for me (I make websites), I’m discovering that there’s a sort of chart of what I should be using it for:

	Computers do it well	Computers find it hard
Humans want to do it	Chess, Go	Emotionally-involved conversations; understanding your new app’s uses
Humans don’t want to do it	Writing boilerplate code; washing dishes	Putting the recycling out

There are two conflicting cells in that table: computers being good at things we want to do ourselves, and computers being no good at something we’re keen not to do ourselves. Let’s put those aside for another time, and concentrate on the two cells we can take advantage of.

We don’t want to do it – let’s automate!

All sorts of machines are available to us these days to take the strain. Washing machines have freed us from the mangle, trains have stopped us from having to walk ten miles to work every day, and mobile phones have obviated the need to communicate with other humans directly.

When you’re boss encourages you to vibe-code the whole of the next project because she thinks that’s how AI coding is building, take a step back and assure them that’s not the case. Copilot and the like will work best when you use it to automate the tedious task of writing out long bits of code (I’m thinking tables, new class files) and to remind you of some syntax you’ve never quite committed to memory.

I’d rather do that myself – get out my way!

The tasks you keep to yourself are the decisions that computers find hard. AI, for example, will tend to reproduce average examples composed of all the designs that have gone before. That creates cliched, and perhaps slightly misguided, designs.

Maybe that’s fine (and maybe that’s what your boss finds acceptable) but if you want to make yourself essential, double down on the things only a human can bring to the party: architecture perfectly suited to the task at hand. Knowing what parts become obsolete and can be thrown away. Knowing what separates a functional app (one with a button for every task) to a usable one (one with controls of the right type arranged in the right way on-screen).

Help is at hand

Cory Doctorow coined the term ‘reverse centaur’, where the human becomes a mere tool of the computer. Doctorow points out that it should be the other way around: the computer should be used to carry out the will of the human. Anything else can be soul-destroying, and will probably spit out sub-par products.

So if you’re wondering about whether you can get a computer (or AI) to improve your productivity without handing over your soul, look for the things it’s good at that you’re glad to be shot of. Make the decisions at the level of detail that you always have, but get the AI to carry them out. And then, most important of all, check that it’s done what you would have done. If it hasn’t, has it done it a better way? If it’s done so much that you can’t tell, maybe you need to break down the task a little.

One day we’ll get it to take the bins out. Until then, get the computer chess out and play a game.