How to Build Taste When AI Makes Everything

For the last couple of weeks my feed has been one long argument about taste. Paul Graham posted that when anyone can make anything, the differentiator becomes what you choose to make. A well-known engineer declared that taste is now the real engineering skill. Then the head of product at Linear pushed back hard: you probably do not have better taste than the AI, so stop pretending that is your edge. Someone else added the line that stuck with me, that the whole taste conversation works precisely because it is vague, impossible to disprove, and flattering to the person invoking it.

All of that is the disposable part. The names will rotate, the quote-tweets will scroll away, and next month the same fight will restart under a different hashtag. What does not go stale is the question underneath it, which is a real one: when a machine can produce a competent version of almost anything in seconds, what exactly is the human contribution, and is “taste” a serious answer or a comfortable dodge?

My honest view is that both camps are half right, and both are missing the same thing. The people selling taste as a moat rarely say what taste is or how you would build more of it, which is why the skeptics can wave it away as ego. And the skeptics are correct that you will not win by having generically better taste than a model that has already absorbed every design award and every bestseller. But they draw the wrong conclusion from a true premise. Taste is real, it is buildable, and the version that matters is not the version they are attacking. This is a piece about the version that matters, and how you build it on purpose.

Table of contents

• What the taste fight is really about
• The anatomy of taste
• What AI actually took from you
• Averaged taste versus authored taste
• How to build each layer
• Taste is a refusal system
• The proof that taste is a moat
• Taste debt: what you ship when you skip the refusal
• The taste map
• What the skeptics get right, and why it does not save them
• What to do Monday morning
• FAQ

What the taste fight is really about

Start with the numbers, because they explain why this argument suddenly feels urgent instead of academic. By 2026 the cost of making a passable thing has fallen to roughly zero. One estimate from Stanford’s Internet Observatory found that individual content farms now push out an average of 94,000 articles a day each. Roughly 34 million AI images are generated daily across the two thousand or so tools that make them. An Ahrefs study put the share of newly created web pages that contain AI-generated text at about 74 percent, and close to 58 percent of new content shows the tells of low-effort machine writing. Google’s content removal actions rose more than 300 percent against a 2023 baseline. The polite word for the result is abundance. The word people actually use is slop.

When production was expensive, being able to produce was itself the edge. You could win by simply shipping more, faster, or at all. That edge is gone. When everyone can generate a clean landing page, a serviceable logo, a grammatically perfect essay, and a working prototype before lunch, the ability to generate stops being scarce and the ability to choose becomes the whole game. That is the real content of Graham’s line. The bottleneck moved from making to selecting.

This is exactly why a company like Taste Labs could raise 18.5 million dollars in mid-2026 on a mission to teach models to stop producing slop, and why its founder defined taste in a way worth borrowing: the bar of quality in the absence of correctness. Most of the important decisions a builder makes have no correct answer. There is no unit test for whether a headline lands, whether an onboarding flow feels trustworthy, or whether a product has a soul. Taste is the faculty you use when correctness runs out, and correctness runs out earlier and more often than engineers like to admit.

So the fight is not really about whether taste matters. Everyone quietly agrees it does. The fight is about whether taste is a thing you can name, build, and be held accountable for, or whether it is a nebulous quality that people with good outcomes claim after the fact to explain their luck. The skeptics are attacking the second version, and they are right to. My job in the rest of this piece is to give you the first version, because a thing you can define is a thing you can train.

The anatomy of taste

Here is the model I actually use. Taste is not one thing, it is four things stacked inside each other, and almost everyone builds only the outermost layer and then wonders why their judgment feels shallow. Think of it as a set of nested refinements, each one narrower and harder to acquire than the one around it.

The outer layer is exposure. It is the raw material, the library of everything you have seen. A person with wide exposure has looked at thousands of products, read across genres, studied work in fields next to their own. Exposure is necessary and it is where most people stop. They confuse “I have seen a lot” with “I have taste,” which is like confusing owning a lot of books with being able to write one.

Inside exposure sits discrimination, the ability to say why one thing is better than another in words, not just a shrug and a feeling. This is the layer that separates a critic from a fan. A fan knows what they like. Someone with discrimination can tell you that this interface feels calmer because it uses two type sizes instead of five, or that this paragraph drags because every sentence is the same length. Feeling gets converted into language, and language is what makes taste teachable and transferable.

Inside discrimination sits a point of view, an actual stance about what is worth doing. Discrimination lets you evaluate anything. A point of view means you are for some things and against others, on purpose, even when the crowd disagrees. This is where taste stops being a receiving skill and becomes a creating one. And at the very center, the smallest and rarest layer, sits refusal: the willingness to act on the point of view by killing work that does not meet it, including your own, including work that is already good. Refusal is the layer that actually shows up in the product, because it is the only one a customer ever sees. The outer three are private. The inner one ships.

What AI actually took from you

Now the uncomfortable part, the part the skeptics are right about. A modern model has more exposure than any human who has ever lived. It has read more, seen more, and can recall more references instantly than you will absorb in a lifetime. It also has strong discrimination, because it has ingested millions of critiques and can explain, fluently, why one design is cleaner than another. So on the two outer layers, exposure and discrimination, the machine wins. It is not close.

This is why “you do not have better taste than the AI” feels true when someone says it. On the dimension most people mean by taste, which is competent good judgment about quality in general, the model has genuinely absorbed the average of everything good that humans have ever made. Ask it for a nice logo and you get a nice logo. Ask it to improve your copy and it will, usually. The floor of quality has risen to meet the machine, and the machine now sits at that floor comfortably.

But look closer at what kind of good the model produces, because the research on this is unusually clear. Large language models exhibit a well-documented pull toward the center of their training distribution, a behavior researchers call mode collapse. Studies through 2025 and 2026, including work presented at ICML on escaping mode collapse and a growing pile of papers on output homogenization, show the same thing from different angles. Repeated samples from a single model converge on the same handful of ideas. Independently trained models converge on each other. As machine-written text floods the web and becomes training data for the next model, the whole system drifts, in the words of one paper, toward bland central tendencies. Researchers have a name for the endpoint: knowledge collapse, the slow narrowing of expressed ideas into an ever-smaller set.

Read that carefully and the real shape of the machine appears. AI is not a taste machine. It is an averaging machine. It gives you the mean of everything good, which is why its output is always competent and almost never surprising, always safe and rarely specific. It regresses to the middle by construction, and it is getting worse at the edges over time, not better, as it trains increasingly on its own homogenized output. The thing it took from you is the value of being generically good. The thing it structurally cannot take is the value of being particular. Those are not the same skill, and the entire taste argument turns on telling them apart.

Averaged taste versus authored taste

So here is the distinction that resolves the fight. There are two things people call taste, and only one of them is worth building now.

Averaged taste is the ability to reliably produce and recognize consensus quality. It is knowing what good looks like in general and hitting that mark. This is genuinely valuable, and it is exactly what the model now has in abundance. If your edge is averaged taste, the skeptics are right about you. You are competing with a machine on the one axis where the machine is strongest, and you will lose, because it can produce the good average faster, cheaper, and at three in the morning.

Authored taste is something else. It is a specific, consistent point of view, applied without flinching, that a mean-reverting system cannot hold. It is not about being better than the average. It is about being reliably off-center in a chosen direction. The whole value comes from the fact that it is not the consensus, which is precisely the region a model is built to avoid. Authored taste is deviation with intent, repeated until it becomes a signature. Where averaged taste asks “is this good,” authored taste asks “is this ours,” and those two questions pull toward opposite regions of the distribution.

Once you see the split, the skeptic’s argument dissolves. Nan Yu is right that you do not have better generic taste than the AI. That is averaged taste, and you were never going to win there. But nobody with real taste was competing there in the first place. The founders who feel un-copyable are not producing a better average. They are producing a particular thing, consistently, in a way the market can recognize and the machine flattens on contact. I made a related argument in my piece on the read-write inversion, where taste functions as the throughput limit on how fast you can accept or reject cheap output. This piece is about the layer underneath that: not how fast you judge, but whether the judgment is generic or yours.

How to build each layer

The good news is that every one of the four layers is trainable, and the training is boringly concrete. It is not a personality you were born with. It is reps, the same kind of deliberate reps I wrote about in the reps problem, just aimed at judgment instead of production. Here is how to build each layer, and the specific way each one fails when you neglect it.

Notice the shape of the failures. They are a ladder of near-misses. Tourist taste has exposure but no discrimination. Fan taste has discrimination but no point of view. Committee taste has a point of view but no refusal, so it dilutes itself back to the average. Passenger taste has all three private layers but never enforces them, so the work that ships is indistinguishable from someone who has no taste at all. Each failure is one layer short of the thing that would have made it particular. The reason authored taste is rare is not that any single layer is hard. It is that you need all four, and the world rewards stopping early.

If you build only one layer this quarter, build the reference library, because it compounds and it is the input to everything above it. The designers I know with the sharpest judgment do not have magic eyes. They have a tagged, searchable archive of thousands of examples and a habit of adding to it every single day. Taste that looks like intuition from the outside is almost always a large private library plus fast retrieval.

Taste is a refusal system

Of the four layers, refusal is the one worth staring at, because it is where taste actually becomes visible and it is the one AI most obviously lacks. There is a line I keep coming back to: taste is not an eye, it is a refusal. A person with taste is not someone who can see beauty everywhere. It is someone who has spent years learning which good-looking options are wrong for this specific thing, and who says no to them out loud, repeatedly, even when saying yes would be easier and nobody would notice.

This reframes taste from a receiving faculty into a subtractive one. You do not demonstrate taste by what you can appreciate. You demonstrate it by what you are willing to throw away. And this is exactly the move a model cannot make, not because it lacks judgment but because it lacks stakes. A model has no cost to producing another variation, no reputation riding on the specific thing that ships, no accumulated identity that some options would betray. It says yes to everything within the bounds of good, because saying yes is free and it has no self to protect. Refusal requires a someone. It requires a position that certain acceptable options would violate.

There is a simple metric hiding in here, and I have started tracking it: the rejection ratio. For a given piece of work, how much did you generate and how much did you keep? A person operating on averaged taste generates one option and ships it, because the option was good and good was the goal. A person operating on authored taste generates ten and kills nine, because eight of them were good and still wrong for this. A low rejection ratio is a warning light. It usually means you accepted the machine’s average because it cleared the quality bar, and quality was never the bar that mattered. The refusal is the work. If you are not throwing much away, you are not applying taste, you are laundering the consensus.

This is also why taste does not survive on autopilot. Refusal is a muscle, and like the judgment reps I described in the reps problem, it weakens fast when a tool starts making the calls for you. Every time you ship the first competent thing the model hands you because it is fine and you are tired, you take one rep away from the faculty that was supposed to be your edge. The tool does not just save you the work of producing. It quietly offers to save you the work of choosing, and choosing was the whole job.

The proof that taste is a moat

The skeptics have a fair challenge to all of this: show me the money. If authored taste is real and buildable and defensible, point to a business that won on it, not a manifesto that celebrates it. Fair. Look at Linear.

Linear built a project management tool in one of the most crowded, least glamorous categories in software, going up against Jira and a dozen well-funded clones. It won, reaching a valuation above a billion dollars on a marketing spend reported in the low tens of thousands of dollars, by treating craft as the entire strategy. Not features, craft. Every interaction targets sub-hundred-millisecond response, with issue updates clocking around 47 milliseconds against Jira’s roughly 3.2 seconds, a difference of nearly seventy times that users feel in their hands even if they never measure it. The team said no to the standard startup playbook of shipping fast and ugly and fixing it later. They held a specific, unfashionable position, that speed and restraint and polish were worth slowing down for, and they enforced it on every screen.

Here is the sentence from their story that should end the taste debate, quoted almost exactly as their followers put it: competitors can copy features, but they cannot copy taste without the team that has it. That is the definition of a moat. A feature is a spec, and a spec can be handed to anyone, including a model, and reproduced. Authored taste is not a spec. It is a thousand small consistent refusals made by a specific group of people who share a point of view, and it does not transfer with a screenshot. Superhuman ran the same play in email, another category everyone thought was finished, charging thirty dollars a month for a faster, more opinionated inbox and holding retention north of eighty percent by obsessing over details competitors thought were beneath them. Neither company had better averaged taste than the market. They had a sharper authored one, and it compounded into a business.

This is the practical answer to “data and distribution are the only real moats.” Those are moats, and I have argued as much elsewhere. But authored taste is upstream of both. It is what makes a product worth talking about, which is how you get organic distribution, and it is what makes people choose you when the feature lists are identical, which is what protects you when the market floods. I wrote about taste as a moat with no expiry clock before this current fight broke out. That piece argued why taste is defensible. This one is about the mechanism the skeptics keep demanding: not that taste is a moat, but which kind of taste, built how.

Taste debt: what you ship when you skip the refusal

There is a cost to skipping the inner layers, and it does not show up right away, which is what makes it dangerous. I call it taste debt, and it is worth separating carefully from a related idea I have written about, cognitive debt. Cognitive debt is about you: the atrophy of your own thinking when you outsource it. Taste debt is about the product: the drift of your output toward the mean when you ship un-refused machine work into it, decision after decision, and it accumulates on the thing customers see whether or not your own faculties decline.

The mechanism is simple compounding. Each time you accept the model’s competent average without running it through your point of view and your refusal, you move one product decision toward the center of the distribution. One such decision is invisible. A hundred of them, across your copy, your interface, your onboarding, your emails, your error messages, and your pricing page, and your product has quietly become the average of every product in your category. It is fine. It clears every quality bar. And it is completely interchangeable, because you built it out of the same consensus every competitor is drawing from, from the same models, trained on the same homogenized web.

This is the real risk of the slop era for a serious builder, and it is not that your product will be bad. Bad you would notice. The risk is that your product will be good in exactly the way ten thousand other products are good, and that the sameness will be invisible from the inside because every individual decision looked reasonable. Taste debt is the gap between “each choice was fine” and “the whole thing has no signature.” You pay it down the only way debt ever gets paid, deliberately, by re-inserting refusal into the loop: taking the machine’s output as a draft rather than a verdict, and asking of each piece not is this good but is this ours, and killing the ones that are merely good. I touched the felt version of this in the velocity illusion, where speed masks the fact that you are not actually deciding anything. Taste debt is the residue that illusion leaves in the product.

The taste map

To make the choices concrete, map two of the layers against each other. On one axis, how much exposure and discrimination you have, which is roughly your ability to tell good from bad. On the other, how authored your point of view is, which is whether you hold a stance or default to consensus. Four positions fall out, and only one of them is defensible against a machine that owns the average.

The bottom-left is derivative work, no eye and no stance, which is the pure slop the tools now produce at industrial scale. The top-left is the confident amateur, all opinion and no reference, the person with strong views and a thin library who is loud and frequently wrong. Most advice about “having a point of view” accidentally pushes people here, into stance without the exposure to back it, which is its own kind of noise.

The bottom-right is the trap that matters, because it is the one talented people fall into. Call it tasteful but generic: a genuinely good eye, real discrimination, and no authored stance. This person makes clean, competent, correct work every time, and it is exactly, precisely, what the AI now hands out for free. If you have spent years developing a good eye and you feel the ground shifting under you, this is why. Your averaged taste was a real achievement and it has just been commoditized. The move is not to get a better eye. It is to add the vertical dimension, to take a position, to become authored. The top-right corner, sharp eye plus real stance applied through refusal, is the only quadrant the averaging machine cannot occupy, because occupying it requires being someone in particular, on purpose. That is the whole target. Everything in this piece is a route to that corner.

What the skeptics get right, and why it does not save them

Let me steelman the other side properly, because they are more right than the taste-is-everything crowd wants to admit. The skeptics make three good points. First, taste as usually invoked is unfalsifiable, a word people reach for to explain their wins without having to specify anything, which makes it a perfect vehicle for ego. Second, you almost certainly do not have better generic judgment than a frontier model, so betting your edge on out-tasting the machine in general is delusional. Third, plenty of people hide behind taste to avoid the harder, more measurable work of distribution, engineering, and sales. All three are true. I would not argue with any of them.

And none of them survive contact with the distinction this piece is built on. The unfalsifiable version is averaged taste dressed up as mystique, and I agree it is mostly ego. But authored taste is falsifiable. It makes a prediction: your work will be recognizably yours, chosen consistently against the consensus, and a stranger will be able to tell your output from a competitor’s with the labels removed. That is testable. Do the blind test. If nobody can pick your work out of a lineup, you do not have authored taste, you have a good eye and no signature, and you should stop calling it taste and start building the stance. The skeptics are right about the vague version and they are attacking a straw man of the real one.

Here is the sharpest way I can put the whole thing. The skeptics say you cannot out-taste the AI, and they are correct, and it does not matter, because being better than the average was never the winning move. The winning move is being reliably off-center in a direction you chose, which is the one region a system built to regress to the mean structurally avoids. You are not competing with the machine on quality. You are competing on specificity, and specificity is the machine’s blind spot by design. The biggest mistake I see talented people make right now is pouring effort into out-qualitying a model that has already won on quality, when the open ground is out-specific-ing a model that cannot afford to be specific about anything. Do not try to be better than the average. Try to be someone the average cannot contain.

The honest counterweight, because there always is one: authored taste is a real risk, not a free lunch. A strong point of view held against a weak library is just being wrong loudly, and a signature that the market does not want is a failed business no matter how consistent it is. Taste does not exempt you from being right about what people value. It is a bet that a specific direction is worth committing to, and specific bets can lose. The skeptics are correct that taste is not a substitute for distribution or judgment about the market. It is the thing that makes your distribution worth having once you have it. Both are true, and a serious builder holds both, which is part of the larger founder operating system for the AI age I have been mapping across these essays.

What to do Monday morning

Enough theory. Here is the concrete work, and all of it fits in a normal week.

Run the taste test on yourself. Pull up the last five things you shipped that a model helped you make. For each one, write a single sentence explaining exactly why you kept it over the alternatives you rejected. If you cannot, because there were no alternatives and you shipped the first competent output, you just found your problem. You have preferences, not taste, and preferences are what the machine already has. Do this for a week and the gap becomes impossible to unsee.

Build the reference library this week. Start a single searchable place and add to it every day. Screenshots, paragraphs, products, whatever is best and worst in and near your field. Tag each one by the principle it demonstrates, not by how it looks, so that in three months you can pull the right example in under a minute. This is the highest-return habit on the list because it feeds every other layer, and I put it near the top of what is worth learning at all in the age of cheap output.

Write your three refusals. Not what you like, what you refuse. Three things that are common, competent, and acceptable in your category that you will not do, on purpose, because they are not yours. These are the spine of an authored point of view. Post them somewhere public so you cannot quietly abandon them when they get inconvenient. A stance you keep private is a preference. A stance you commit to is taste.

Track your rejection ratio for one project. Count how much you generate and how much you keep. If you are keeping almost everything, you are laundering the average, and you should deliberately generate more options and kill more of them until the throwing-away feels like the real work, because it is. Refusal is the layer that ships.

Run the blind test once a quarter. Strip your branding off a piece of your work, mix it with three competitors’ and one raw model output, and ask someone honest to pick yours. If they can, your signature is real and getting stronger. If they cannot, you are in the tasteful-but-generic corner, and you now know exactly which layer to build next. This is the closest thing to a unit test that judgment allows, and for a solo builder it is one of the few faculties worth protecting as an incompressible part of the work no tool can absorb.

None of this requires talent you were born with. It requires reps aimed at judgment instead of production, a library, a stance, and the willingness to throw good work away. The machine gives you the average of everything good, instantly and for free. Your entire job now is to be specific on purpose, because specificity is the one thing a system that regresses to the mean cannot copy. Build the four layers, keep the refusal muscle alive, and you own the one corner of the map the flood cannot reach. If you want the wider context this sits inside, the AI opportunity map lays out where judgment like this pays off across a business.

FAQ

What does it mean to build taste when AI can make everything? It means deliberately developing four stacked faculties: exposure (a searchable library of great and weak work), discrimination (naming in words why one thing beats another), a point of view (a held stance on what is worth making), and refusal (killing work that does not match that stance). AI has flooded the outer layers by producing competent, average work at zero cost, so the human edge moves inward, to holding a specific and consistent point of view that a mean-reverting model cannot reproduce.

Is taste actually a skill you can build, or are you born with it? It is a buildable skill, and treating it as innate is the main reason people never develop it. Taste is compressed judgment learned from exposure, practice, and reps, the same way any expert intuition is built. Designers with the sharpest eye almost always maintain a large tagged reference library and a daily habit of analyzing one strong and one weak example. The intuition that looks like magic from the outside is usually a big private library plus fast retrieval and a committed stance.

Do humans really have better taste than AI? Not on average, and that is the wrong frame. A frontier model has more exposure and stronger general discrimination than any individual, so on generic quality it wins. But models regress to the mean by construction, a behavior researchers call mode collapse, and they get more homogenized over time as they train on their own output. So humans do not beat AI at average quality. They can hold something AI structurally cannot: a specific, off-center, authored point of view applied consistently.

What is the difference between averaged taste and authored taste? Averaged taste is the ability to produce and recognize consensus quality, hitting the mark of what good looks like in general. That is exactly what AI now does cheaply, so competing there is a losing game. Authored taste is a specific, consistent point of view held against the consensus, deviation with intent repeated until it becomes a signature. The value comes from being recognizably particular, which is the one region a mean-reverting model is built to avoid.

Why is AI-generated work so often bland even when it is technically good? Because a model produces the statistical center of its training data. Research on mode collapse and output homogenization through 2025 and 2026 shows models converging on the same ideas, both across repeated samples and across independently trained systems, and drifting toward bland central tendencies as AI content becomes training data for the next model. The output is competent because it is the average of good work, and forgettable for the same reason: the average of everything is specific to nothing.

What is a refusal system, and why does it matter more than a good eye? Taste shows up not in what you can appreciate but in what you are willing to throw away. A refusal system is the practice of setting a standard from your point of view and killing work that does not meet it, including work that is good but not yours. It matters because refusal is the only layer a customer ever sees, and it is the one a model most obviously lacks, since a model has no stakes, no reputation, and no self that certain options would betray. Track your rejection ratio: if you keep almost everything you generate, you are shipping the average, not applying taste.

Can taste really be a business moat? Yes, and the clearest proof is a company like Linear, which won a crowded software category on craft, reaching a valuation over a billion dollars on marketing spend in the low tens of thousands, holding a specific unfashionable stance about speed and polish and enforcing it everywhere. Competitors can copy features, which are specs anyone can reproduce, but they cannot copy authored taste without the team that holds it, because it is a thousand consistent refusals rather than a document. Superhuman ran the same play in email. Taste is upstream of distribution and pricing power both.

What is taste debt? Taste debt is the drift of your product toward the mean when you repeatedly ship un-refused machine output. Each accepted average is one decision moved toward the center of the distribution, and across a hundred small choices your product becomes interchangeable with every competitor drawing from the same models and the same homogenized web. It is different from cognitive debt, which is the atrophy of your own thinking; taste debt lives in the product, not your head. You pay it down by treating model output as a draft, asking of each piece not is this good but is this ours, and killing the ones that are merely good.