Taste is the skill AI cannot hand you

In 2024, on an Alphabet earnings call, Sundar Pichai said that more than a quarter of all new code at Google is generated by AI, then reviewed and accepted by engineers. Sit with the second half of that sentence. The machine writes the code. A person decides whether it is any good. The work that used to define a senior engineer, the typing, the syntax, the getting-it-to-compile, has been handed to a tool. What is left for the human is the part that was always the hardest to teach: judgment about what to keep.

That split is showing up everywhere at once. A founder can spin up a landing page, a pitch deck, a cold email sequence, and a working prototype in an afternoon. The cost of producing a thing has fallen through the floor. And the moment production gets cheap, the thing that used to be a side quality, knowing what is worth producing and whether the output is actually good, becomes the whole game. I call that taste. Not taste as in which font looks nicer. Taste as in the trained ability to look at a piece of work and know, quickly, whether it is right.

Here is my claim, and I will defend it for the rest of this piece: taste is the one skill AI cannot hand you. It will write your code, draft your copy, and summarize your research. It will not tell you whether any of it deserved to exist. That judgment is yours to build, it compounds like nothing else you own, and the casual way most people use AI is quietly eroding it.

The bottleneck moved, and most people missed it

For most of the history of knowledge work, the constraint was production. You had an idea, and then you spent weeks turning it into a document, a model, a feature, a campaign. The idea was cheap and the execution was expensive. So we built our whole sense of value around execution. We hired for it, we paid for it, we measured people by how much they could make.

Generative AI inverted that. Making things is now close to free, and the scarce input is knowing which things to make. You can see the consequences in the enterprise data. In 2025, a widely discussed MIT report titled “The GenAI Divide: State of AI in Business” found that roughly 95 percent of corporate generative AI pilots were failing to produce meaningful financial returns, with only about 5 percent achieving rapid revenue growth. The models in those failed pilots were not broken. The same models were working fine for the 5 percent. What separated the winners was not better technology. It was better judgment about where to point the technology and how to tell whether it was working.

This is the quiet shift behind the noise. When everyone has the same tools, the tools stop being the advantage. I wrote a whole piece on why the hard part of an AI product is the evaluation, not the model, and the personal version of that argument is the same: your edge is no longer what you can produce, because the machine produces too. Your edge is the quality of your “yes” and your “no.” That is a skill. It is learnable. And almost no one is deliberately training it, because we are all too busy being amazed at how much we can now make.

What taste actually is, stripped of the mystique

People treat taste like a personality trait you either have or you do not. I think that is lazy and mostly wrong. Taste is two concrete abilities working together, and both of them are trainable.

The first is knowing what is worth doing. Out of a hundred things you could build, write, or chase, which small set actually matters? This is a filtering skill. It is the ability to look at a long list of plausible options and feel, fast, which ones are real and which ones are noise dressed up as opportunity. The second is knowing whether a finished thing is good. You look at the output, the draft, the design, the plan, and you can tell where it is strong, where it is fake, where it will break under load. This is a discrimination skill. It is pattern recognition built from having seen a great many examples, good and bad, and having paid attention to the difference.

Notice that neither ability is about making the thing. Both sit on either side of production, the choosing before and the judging after, and those are exactly the two places AI leaves empty. The machine happily makes whatever you point it at, as much of it as you want, with no opinion about whether the target was worth hitting or whether the result is worth keeping. The chart below is how I picture the shift.

The lines cross, and the space between them after the crossing is where the real gains now live. If you keep investing only in production, you are riding the falling line down. If you invest in judgment, you are climbing the rising one. My bet is that the people who compound over the next decade are the ones who notice this early and act on it on purpose, the way earlier generations noticed that distribution, not product, was the real constraint.

You build taste the same way you always did, with more reps

The good news, and the uncomfortable news, is that building taste is not mysterious. It works the way skill acquisition has always worked. In 1993, the psychologist Anders Ericsson and his colleagues published the research that put deliberate practice on the map: expert performance comes not from raw talent or hours logged on autopilot, but from focused repetition against immediate feedback, repeated at the edge of what you can already do. You try, you see the result, you adjust, you try again. Thousands of times.

Taste is built the same way, and the feedback loop has three parts. You expose yourself to a large volume of examples. You make your own judgments about them, out loud or on paper, before you see the verdict. Then you check your call against reality and against people whose judgment you trust. Do that enough and your gut starts to encode patterns you could never articulate. A great editor does not run a checklist. She reads a sentence and winces. That wince is ten thousand sentences of compressed feedback firing at once.

Three things actually move the needle here, and I would rank them in this order. First, study the best work in your domain closely enough to see why it is good, not just that it is. Most people consume great work as entertainment and learn nothing; the reps come from taking it apart. Second, produce your own work and get real feedback on it, from the market or from people who will tell you the truth. Third, write down your predictions before you learn the outcome, because a judgment you never committed to is a judgment you cannot grade. I keep a running note of calls I made and how they turned out, and it is humbling in the specific way that makes you better.

None of this is fast. That is the point. The learning system I use to get sharp in a new domain is really just this loop run at high intensity: heavy exposure, forced judgments, fast feedback. The reason taste is defensible, the reason it compounds, is precisely that it cannot be downloaded. You cannot prompt your way to a trained eye.

The crutch problem, or how AI can quietly cost you the thing that matters

Here is where it gets dangerous, and where I think most people are sleepwalking. The same feedback loop that builds taste is the loop AI invites you to skip.

In early 2025, researchers from Microsoft Research and Carnegie Mellon University published a study called “The Impact of Generative AI on Critical Thinking,” surveying 319 knowledge workers who use these tools at least weekly. The finding that stuck with me: the more a worker trusted the AI, the less critical thinking they applied to its output. And people who were less confident in a task were the most likely to accept the AI answer wholesale, without scrutiny. The authors described the shift as a move from gathering information to verifying it, and from executing tasks to overseeing them. They warned of cognitive offloading and a possible atrophy of skills we might want to keep.

Read that against everything I just said about how taste is built. Taste comes from making your own judgments before you see the answer. AI is a machine that hands you the answer first. If you let it, every hard call gets outsourced at exactly the moment the call would have taught you something. You stop generating your own hypotheses about what good looks like, because the tool always has one ready. The reps disappear. And here is the trap inside the trap: the study found the people most likely to offload were the ones who felt least capable, which means the tool erodes judgment fastest in the people who have the least of it to start with. The rich get richer and the uncertain get hollowed out.

I am not anti-AI. I build AI-native companies and I use these tools all day. But I treat the convenience as a known hazard, the way a strong swimmer respects a riptide. The danger is not that AI is bad at the work. The danger is that it is good enough to make you stop practicing the one thing it cannot do for you.

How I use AI without handing it my judgment

So what does it look like to use these tools hard and still keep your eye sharp? My rule is simple to say and annoying to follow: delegate the production, never the decision. The machine can make the thing. I keep the choosing and the judging for myself, and I do them before and after the machine runs, not instead of it.

In practice that means I form my own view first. Before I ask a model to draft the strategy memo, I write down, in ugly bullet points, what I think the answer is and why. Then I let it draft. Then I compare. The gap between my take and its take is where the learning happens, and I would lose that entirely if I prompted first and thought second. When the output comes back, I read it like an editor who assumes something is wrong, not like a reader hoping it is right. I am hunting for the fake sentence, the number with no source, the confident paragraph that would collapse under one real question.

The difference between the two postures is the difference between taste you own and taste you rent. The table makes it concrete.

The left column is faster this week. The right column is the only one that is still paying you in three years. It also means grading yourself by a standard you set rather than by applause, which is the same muscle behind keeping an inner scorecard. This is the same logic I apply to making decisions when you cannot know the answer in advance: the quality of the process is the only thing you control, so you protect it even when a shortcut is sitting right there.

Taste without the ability to ship is just a loud opinion

I should argue against myself, because there is a real counter-case and I do not fully escape it. The honest objection is that taste, celebrated on its own, becomes an excuse. The world is full of people with exquisite judgment who never make anything, critics who can tell you precisely why every attempt is flawed and have never been in the arena themselves. If production is now cheap, the risk is that we romanticize the judging and quietly let the doing rot. A person who can only evaluate is not valuable. They are a comment section.

There is a second objection, sharper for founders. In a brand new market, speed often beats taste. The first person to ship a rough version and put it in front of real users learns things that no amount of refined judgment could have told them from the armchair. Early on, taste can even mislead you, because your trained eye was trained on the old world and this is a new one. Sometimes the ugly, fast, slightly embarrassing thing is correct, and the tasteful restraint is just fear wearing a nice coat.

I take both of these seriously, and here is how I hold them. Taste and execution are not rivals; they are a loop. Judgment tells you what to ship and how to read the result. Shipping gives judgment the feedback it needs to improve. The failure mode is picking one and worshipping it. The critic who never ships has taste that slowly goes stale from lack of contact with reality. The founder who ships endlessly without ever sharpening their read just produces more noise, faster. What AI changes is only the mix: since it now handles so much of the doing, the scarce input you have to supply on purpose is the judging. That does not make shipping optional. It makes the judging the part you can no longer afford to skip.

Train your eye before you train the model

If you buy even half of this, the move is to start treating judgment as a thing you practice, not a thing you wait to receive. A few concrete places to begin.

Pick one domain where your taste actually matters for your work, and spend real time with the best examples in it, not scrolling past them but taking them apart until you can say why they land. Start a predictions note and write down your calls before outcomes arrive, so you can grade your own judgment honestly instead of remembering yourself as right. Change how you use your AI tools: form your view first, let the model draft second, and read everything it gives you as a skeptic hunting for the weak joint. When a task feels hard and you feel the pull to just accept whatever the model says, treat that pull as the signal to slow down, because that is the exact moment the rep is worth the most. And protect a little daily friction on purpose, some work you still do by hand, because a muscle you never use does not stay strong out of loyalty.

The builders I would bet on over the next decade are not the ones with the best access to models, since everyone will have that. They are the ones who used the cheap production to free up time, and then spent that time deliberately sharpening the one thing the models hand back to them untouched: the judgment about what is worth making and whether it is any good. The tools will keep getting better at the doing. Make sure you are getting better at the deciding. If you want the fuller version of how I think about running your own mind as a founder, it lives in my founder operating system, and the broader map of where I think the real opportunities sit is in the AI-native founder playbook.

Questions founders ask me

Is taste just a fancy word for experience

Not quite, and the difference matters. Experience is time spent. Taste is pattern recognition extracted from that time, and you can have a lot of one without much of the other. I know people with twenty years in a field whose judgment never improved, because they repeated year one twenty times without ever grading their own calls. Taste is what you get when experience runs through a feedback loop: a judgment made, an outcome observed, the gap studied. You can also build it faster than raw experience would suggest by being deliberate, studying the best work closely and forcing yourself to predict before you peek.

How do I build taste without slowing myself down

You accept that the first months are slower and the payoff comes later, the same trade any skill demands. But the slowdown is smaller than people fear, because the reps piggyback on work you are already doing. You are going to write that memo anyway, so write your own take first before you prompt. You are going to review that output anyway, so review it like a skeptic. The only real added cost is the predictions note, which takes minutes. Speed and taste feel opposed in the moment and compound together over time.

Can AI help me build taste rather than erode it

Yes, if you invert how you use it. Instead of asking the model for the answer, ask it to critique yours. Make your call first, then have it argue the other side, list what you might be missing, or show you three examples of how experts handled the same problem. Used that way it becomes a sparring partner that sharpens your judgment instead of a vending machine that replaces it. The deciding factor is always the order: your view first, the machine second.

What if my field changes so fast that old taste is worthless

Some of your specific patterns will go stale, and you should expect that. But the meta-skill, the habit of forming judgments and grading them against reality, transfers across every change. The person who built real taste in one domain learns the next one far faster, because they already own the loop that builds it. What goes out of date is the content of your judgment. The engine that produces judgment does not.

How do I know if my taste is actually any good

Grade it. This is why the predictions note matters so much: it turns a vague feeling of being right into a record you cannot argue with. Track the calls you make, check them against what actually happened, and look for a pattern of being more right than the people around you over time, not in any single instance. Good taste shows up as a track record, not as confidence. If you are unwilling to write your predictions down, that reluctance is usually a sign the confidence is not yet earned.

Where should a founder with limited time start

Pick the one or two decisions that actually move your company and pour your judgment there, instead of spreading a thin layer of attention over everything. For a founder that usually means who you hire, what you build next, and who you sell to first. Let AI take as much of the production around those decisions as it can, and spend the time it frees on getting the decisions themselves right. Taste is not a general polish you apply everywhere. It is a scarce resource you concentrate on the few calls that compound.

I’m a four-time founder building AI-native companies. I write about the founder mindset, AI opportunities, and the mechanics of building with AI at vikasmalpani.com.