TasteIs aTechnology.

For most of the history of design, the hard part was making something. An idea cost almost nothing, and making it real cost days or weeks. A logo took a week of drawing. A poster took a printer. Because production was expensive, the world contained relatively few finished things, and a great deal of the craft of a designer lay in the hands.

That economy has ended. A person with a subscription can now produce a hundred logos before lunch, along with the brand story, the campaign and the website. Most of them are competent. Many of them are pleasant. All of them arrive with no effort, and more or less all of them look like one another.

I use these tools daily, and I am not here to mourn. I build with them. In my view the arrival of cheap generation is one of the most interesting things to have happened to my profession. But I think most people have misunderstood what it means. They ask whether machines will replace designers. I think that is the wrong question. The better question is this: once making is nearly free, what becomes valuable?

My answer is taste. And I want to argue something stronger than the usual remark that taste matters. I want to argue that taste is a technology.

Once making is nearly free, what becomes valuable?

Look at the two lines. As the cost of producing an option falls, the number of options explodes. And as the number of options explodes, the value of choosing among them rises. A library with ten books needs no librarian. A library with ten billion books needs nothing else. We have entered the age of the ten billion books, across every visual and written form.

When people call taste a matter of opinion, they mean that it cannot be discussed. I think that belief has done great damage. It lets bad work escape judgement, and it leaves good practitioners unable to explain what they do. So let me make the case that taste is not mysterious at all. It is a compressed form of judgement with a structure, a training regime and an economic function. That makes it a technology, and perhaps the most important one that a person can personally own.

Watch an experienced art director look at a screen. In about a second, they say: no, that one. Ask why, and they may struggle. To an outsider this looks like magic, or arrogance. I think it is neither. What you are seeing is compression.

Underneath that glance lies a long stack. There are thousands of examples seen and remembered, compressed into pattern. There are principles, the things that can be said aloud: this is too symmetrical, that line has no reason to be there. There is history, the knowledge of what has been done before, so that an idea which seems fresh to a beginner is recognised as a cliché of 1994. And at the base lie values: a view about what the work is for, and for whom.

Herbert Simon and others who studied expertise found something similar in chess masters. A grandmaster's quick intuition turned out to be pattern recognition built up over many thousands of hours. The quick judgement is real, and it was bought at a high price. Daniel Kahneman's distinction between fast and slow thinking helps here too: a fast judgement can be a trained one, provided that it has been built by long experience with honest feedback.

This is why taste is compressed judgement. It stores years of slow thinking in a form that can be used in one second. That is precisely what a good technology does. A compression algorithm allows a large thing to be moved quickly. Taste allows a very large amount of experience to be applied to a decision, at the speed of a glance.

Taste stores years of slow thinking in a form that can be used in one second.

So where does it come from? It is not inherited. It comes from a loop.

You see a great deal of the best work, which means leaving your own field and your own decade. You make things with your own hands, and you fail. You compare: this is the most underrated exercise of the lot, simply holding two things next to each other until you can feel the difference. You name that difference out loud, in words, since an unnamed feeling cannot be taught or defended. You refuse things that are merely fine. Then you put the work into the world and you live with the consequences.

Naming and refusing are the two steps that most people skip. Plenty of people look at lots of work. Far fewer force themselves to state exactly why one thing is better. And refusal is painful, particularly when a client is happy, the deadline is close and the work is acceptable. But each refusal strengthens the faculty. Each time you let something pass that you knew was only fine, it weakens a little.

There is a danger here that deserves attention. If young people no longer make things by hand, because a machine can do it instantly, then where will the next generation's taste come from? You cannot acquire judgement about drawing without having drawn badly for years. I worry about that, and I do not have a complete answer. I would want a young designer to do things the slow way first, on purpose, even when a faster way exists. The slow way is how the compression is built.

Why does so much machine-generated work look the same? There is a technical reason, and it is worth understanding.

A generative model learns the statistical shape of a huge body of existing work, and its most probable output sits near the centre of that distribution. In plain words, its first answer tends towards the average of everything. Ask for a modern brand and you will receive the modal modern brand. It is competent, and it is nobody.

This is where the word slop has come from. It refers to generated output produced without judgement, at volume. You will recognise the signs. The purple-blue gradient. The earnest stock optimism. The sentence that sounds profound for one second and means nothing. The rhythm of a paragraph that is too tidy. The same three-part list, again.

In the BRVND guideline this is written down as a formal position: the Anti-AI-Slop Doctrine, which calls for the ruthless elimination of synthetic clichés, muddy greys, purple-blue neon gradients and superficial marketing buzzwords. It is an operating rule for a brand system that itself uses AI heavily. This is not a contradiction. It is the entire point. A system that generates at machine speed needs a stronger immune system than a human studio has ever needed.

A system that generates at machine speed needs a stronger immune system than any studio has ever needed.

The table is meant to be used and not admired. Notice that it does not say to avoid AI. It says to avoid the unconsidered. The right-hand column gives the reasons, and I think those matter most, since a rule without a reason decays into a superstition.

The deep point is about distinctiveness. A brand exists to be recognised. Its whole economic function depends on being different from its neighbours in a way that can be remembered. So a tool whose natural tendency is towards the average is, by default, working against the purpose of a brand. If every company takes the first answer, then every company will come to look alike, and brand as a concept begins to dissolve. Generic is the most expensive thing a company can be, since it is paying to be forgotten.

That leaves an enormous opening for anyone willing to do the work. When most output converges on the mean, a small deviation becomes highly visible. A real point of view, a strange colour, a typeface with a history, a sentence that a person actually meant: these stand out against the grey more than ever. In a world of averages, specificity is a superpower. The way to get it is to begin somewhere the model cannot: with this people, this history, this place, this particular problem.

Now let me widen the frame, because this goes beyond design.

If you accept that the supply of content is becoming unlimited, then the most important institutions of the next decades will be those that select. I use the word institutions on purpose. Selection used to be done by editors, publishers, curators, juries, critics and teachers. Those roles have been under attack for thirty years. The internet told us that we no longer needed gatekeepers. Everyone could publish, and the crowd would sort it out.

That was a generous idea, and in some ways it was right. But it assumed a human scale of production. It did not anticipate an unlimited supply of plausible material, some of it produced with the specific intention of misleading. In that environment, a trusted selection stops being a luxury. It becomes infrastructure, in the same sense in which I have argued that visibility is infrastructure. A society needs reliable ways of knowing what is worth its attention, just as it needs clean water.

A trusted selection stops being a luxury. It becomes infrastructure.

It explains why I say curated, not crowded, on Kurd Artist, and why I think a prize with an honest jury has become more valuable than it was. Each of those is a human act of selection with a name attached. In a sense, a good editor is a compression algorithm for a whole culture. They read everything so that you do not have to, and they stake their name on the result.

The word stake matters. What separates taste from a recommendation engine is that a person with taste can be wrong in public and suffer for it. They have a reputation that is exposed by every choice. A model has none. It will recommend anything without blinking, and nothing happens to it if the recommendation was poor. Accountability is part of what makes judgement trustworthy. I do not expect that to change, even as the models improve.

Machines will, of course, get better at imitating selection. You can already ask a model which of ten options is strongest, and sometimes it gives a decent answer. It can be a useful second opinion. Yet its taste is the average of the taste in its training data, and so it will tend to prefer what is already preferred. It will rarely choose the thing that looks wrong today and turns out to be right. That leap, away from the consensus, is exactly what great taste has always done. Nearly every important movement in art looked like a mistake at first.

What does all of this mean for how a creative team should actually work?

The schedule above is a model of what I am moving towards. The most striking thing about it is the size of the generation bar. What used to occupy most of a project now takes hours. Most studios, when they get that time back, simply do more of the same, faster. They produce more options, more rounds and more decks. I think that is a mistake. It multiplies the mediocre.

The freed time should go to both ends of the process. At the front, it goes to understanding: real time with the people, the history and the actual problem. No model can do that for you, because the information is not in its training data. It lives in a conversation in somebody's office, or in the way a founder's voice changes when they talk about one particular thing. Then you decide on an idea, a position in words, before any image exists. If you generate before you decide, the machine will decide for you, and it will choose the average.

If you generate before you decide, the machine will decide for you.

At the back end, the time goes to selection and craft. Generate broadly and quickly, knowing that nearly all of it will be thrown away. Then compare, name and refuse. Finally, take the survivor and finish it by hand. That last ten percent is where the work becomes yours: the correction of a curve, the tightening of a line, the removal of the one element that was there out of habit. A viewer rarely knows what was done. They feel only that it is right.

The roles change as well. The most valuable person in a future studio will not be the fastest maker. It will be whoever can look at two hundred options, say which one is right, and explain why in a way that a client can believe. It is the role of the art director, the editor and the creative director, made central. Hiring has to change accordingly. The useful test is less about software skills and more about what a person can say about a piece of work. Show someone two things and ask which is better and why. Their answer tells you nearly everything.

It also changes what we should be charging for. Clients who see that a logo can be generated in seconds will ask why they should pay for one. The honest answer is that they never were paying for the drawing. They were paying for the judgement: for the thousand wrong answers that were not shown to them, and for someone's name on the right one.

I want to finish with a claim that may sound exaggerated, and I mean it.

Taste is not only an aesthetic faculty. At its depth, it is a moral one. To have taste is to refuse to be satisfied by the merely adequate. It is to care about something that most people will never notice: a margin, a word, the fit of two things. It is to believe that the world is owed some effort. The opposite of taste is not bad taste. It is indifference: the shrug that says it is good enough, and nobody will notice.

Slop is a form of indifference, produced at industrial scale. Every piece of generated filler says to its reader: I did not think that you were worth a person's attention. People feel this, even when they can't name it. There is a small insult in it. As the volume of such material rises, I expect that its opposite will come to feel almost like a courtesy. Something made with care will read as an act of respect.

There is a further stake in my part of the world. Our visual cultures, Kurdish and others, are poorly represented in what machines have learned. Left to the default, a generated Kurdish poster will be a costume, a pastiche of carpets and mountains. Those of us who actually know these traditions have a responsibility to do better than the average, because the average of us does not yet exist, or exists in a form written by others. Taste, here, is a form of self-defence. It is how a culture refuses to be flattened.

So I would give a young designer this advice. Use every tool that is available, and do not be sentimental about it. But spend your life training the one faculty that cannot be downloaded. Look at more than you think you need to. Make things badly. Compare. Say why. Refuse. Learn what has been done, so that you can recognise what has not. Above all, care more than is reasonable about things that hardly anybody will notice.

The machines will keep getting better at making. That is fine. Somebody will still have to know which one is right, and be willing to stand behind it. That is the job now. It may be what the job always was.