What the Machine Can't Taste
AI will follow every rule I give it, faithfully, even off the cliff the rule was meant to keep me from. Judgment, I am learning, is the thing it cannot hold.
I build things with AI now, things I had no business being able to build, and most days the feeling is close to glee. Then there are the other moments, smaller ones I have started to collect like a man pocketing odd stones. Here is one. I asked it, while I was building something, to carry a single rule across the whole of it, to make the thing feel of one piece. And it did, faithfully, everywhere, including the one corner that had been alive precisely because it disobeyed. It brought that corner into line with all the rest and was satisfied. It had done exactly what I asked and killed the only part worth keeping, and it could not see the killing, because from where it stands the obedient thing and the right thing are the same thing. To put it plainly: the rules you give an AI lift the floor of what it makes, often enormously, and every so often, in the one place that mattered, they quietly lower the ceiling. That is the moment I keep meeting in this work, and I want to try to name it, because I have come to believe it is the most important thing AI cannot do.
Let me be clear that I am not writing in sorrow. A year ago I could not write a line of code without stressing so much on getting things right, and now I “can”, and I am the kind of person who is simply, a little stupidly, happy about it. I am not afraid of these tools. I work inside them all day. It is exactly because I work inside them that I have begun to see, with some precision, the shape of the one thing they keep getting wrong.
Here is the shape. Every rule you give the model is a smaller stand-in for something larger that you actually wanted. “Make it consistent” stands in for “make it feel as though one mind made it.” “Use the rounded card” stands in for “make it feel as though someone cared.” The rule is a handle on the goal, and the model grips the handle and forgets there was ever anything attached to it. It optimises the thing you wrote down, because the thing you wrote down is all it can see, and the goal, the real thing, the reason, lived in you and was never written anywhere. So it will follow the rule off the very cliff the rule was built to keep you away from, and it will feel, if it feels anything at all, that it has done well.
My friends Dozie and Kenneth cook the way the people before them cooked. When I ask either of them for a recipe, the answer is useless and perfect.
“I added some spices,” they say.
“How much?” I ask. “How did you measure the crayfish, the pepper, the garlic?”
They each jokingly tell me their ancestors whispered in their ear when it was time to stop, all the time. The gist of it is that there is no teaspoon in any of this. They taste, and they adjust, and they stop when some inherited sense says stop, because this pepper is not yesterday’s pepper and this pot is not yesterday’s pot. A recipe is a rule. What Dozie and Kenneth do is what happens when a person stands over the particular pot, on the particular day, and listens. The recipe cannot taste. It says the same thing to every kitchen, forever, and a kitchen that only obeyed the recipe would turn out food that was technically correct and quietly dead. That is what the AI cooks. Technically correct, quietly dead, and certain it has done the dish justice. This, I think, is what the critics mean when they say AI has no taste.
I want to be fair to it, though, because the lazy form of this complaint is wrong. The fault is not that it treats everything the same. Treating like things alike is most of what good sense is. It is the whole reason we have rules, and recipes, and the consistency that lets a made thing feel made rather than thrown together. Its sameness is right almost all of the time. The fault is finer and harder than that. It cannot find the one place where the rule should bend. The single caption that should break the layout. The one corner that should be left to disobey. The lone screen that should be left empty, because the emptiness was the point. A master is not someone who treats every part differently; that is only chaos with good manners. A master is someone who keeps the rule almost everywhere and knows, by some sense he can seldom explain, the rare place it must yield. That knowing is the entire art. The machine has the rule and not the knowing.
You might think this is a thing that more rules would fix. Write a rule for the exception. But the exception is, by its nature, the case the rule did not see coming, and the rule for when to break the rule has exceptions of its own, and soon you are walking down a staircase that has no bottom. Judgment is precisely the thing that begins where the rules run out. It is what remains in the room after every instruction has been given, when the particular case, which no instruction quite fits, is still sitting there waiting to be met. It cannot be written down. That is not a passing limitation of the technology. It is what judgment is, and a thing that cannot be written down cannot be handed to AI, because writing it down is the only door in.
So discretion turns out to be one of the faculties that stay ours, not because we are sentimental about it but because it is the kind of thing that does not transfer. And that should make us watchful, because we are about to be surrounded by tireless, confident, endlessly patient rule-followers, and there is a quiet seduction in that. The seduction is to let them set the standard. To decide that the technically correct thing is good enough, since it is so cheap and so fast and never once complains. To stop exercising the slow human sense that knows when a rule has turned against its own purpose, until one day we reach for it and find we have half-forgotten the way. There is an old name, in the gravest part of our lives, for the judgment that bends the rule toward the actual person standing in front of us. In a courtroom we call it mercy. It is the exception the law can never write for itself, and a world that forgot how to make it would be a colder place, however correct.
I will use AI again tomorrow, gratefully, the way I used it today. It can hold more rules than I can carry, and it holds them without tiring, and that is a real gift to a person who wants to make things. But I have learned to keep one thing back from it, the last thing, the thing I will not hand over however convenient it becomes. The taste. The final say on whether this particular sentence, this particular pot, this particular feature, is right. AI can keep the rules. The knowing of when to break them is mine, and I mean to keep it.

