Taste is a bet on a standard that doesn't exist yet

Two weeks ago I published a note arguing that taste is domain judgment wearing a nicer word. A designer calls it taste, a consultant calls it judgment, and they’re closer to the same skill than either would admit.

I still think that’s right about who has it. I think it’s wrong about what it is.

The correction came from trying to pull the two words apart and finding that I could. Judgment is being able to define whether something is good or not as per the society’s standard. The standard already exists. You’ve internalized it, and you apply it faster and more accurately than the rest of the room can. Taste doesn’t work like that. With taste there’s no really wrong answer at the moment you make the call. It’s something that in a longer term people will gradually find out — okay, this is the way to go.

Same activity, different target. Judgment answers to a standard that exists. Taste answers to one that hasn’t formed yet.

The moat argument swaps them mid-sentence

Once you hold the two apart, the popular version of the claim starts to wobble.

“Taste is the moat in the AI era” has been talked about too many times. People just say taste, taste, taste, and they don’t seem to define what taste is. When they do reach for evidence, the evidence is almost always about judgment: the senior person who knows which analysis this board actually cares about, which recommendation is technically right but dead on arrival, which slide to cut. That’s real, and it’s scarce, and it’s a standard that already exists. The board has it, and the senior person has learned to predict it.

Then the conclusion arrives dressed as taste, which is the harder and more romantic thing. The argument earns its keep on judgment and cashes out on taste. That swap is why the phrase feels both obviously true and strangely empty.

The distinction also decides how defensible any of it is. A standard that exists can be studied. Someone can sit in enough board meetings, or read enough of them, and get good at predicting the room. That’s a moat made of time and access, and moats like that erode when the time and access get cheaper. A bet on a standard that doesn’t exist yet can’t be studied the same way, because there’s nothing to study yet. Whatever protection taste offers, it comes from that gap, not from being generally hard.

How much of each is in your job

Ask what taste means for a consultant versus a designer and the honest answer is: it depends on the domain.

That sounds like a dodge, but the mix is the interesting part. Most consulting work is judgment — reading a standard that already exists in the client’s head, in the industry’s norms, in what the last three boards approved. The taste fraction is small but it’s the part nobody can staff around: choosing to argue something the client hasn’t asked for and isn’t ready to hear, and being right about it eighteen months later.

Design inverts the ratio. A lot of design is still judgment against known standards — accessibility, convention, the thing users have already learned. But the fraction that decides whether the work matters is the bet: this is what the category should look like next, and there’s no way to check today.

Engineering mostly lives in the judgment column, whatever the word “taste” gets used for. Restraint, knowing what to leave out, refusing the clever abstraction: these get validated by standards that already exist, and fairly quickly. Code that ages badly tells you within a year. That’s a shorter feedback loop than taste gets.

What you’d actually look for in a person

If you’re hiring and the person hasn’t produced anything yet, the thing worth probing is awareness. Awareness of oneself, and the awareness or understanding of a specific population’s reaction to things, and knowing what actions we need to take that can incur or resolve certain reactions.

That’s vague, and it’s vague in a way I don’t think can be fixed by trying harder to define it. It’s a description of someone running a model of other people accurately enough to predict which move produces which reaction. You can test that in conversation more easily than you can test it on a portfolio. Ask someone why a piece of work failed and listen for whether the answer is about the work or about the people it landed on.

The uncomfortable part

Here’s where I’d argue against my own side. Taste is probably more reproducible than the moat crowd needs it to be.

It’s almost like a certain type of pattern recognition that people can’t summarize yet. That’s a different claim from “it’s ineffable.” Unsummarizable and uncomputable are different properties. The person holding it can’t write the rule down, which is the normal condition for skills learned by exposure rather than instruction.

What makes it hard is not the modeling. It’s that the signal is so subtle that collecting the data is the actual problem: you’d need the calls a specific person made, the options they rejected, and the reason, at a volume nobody records. That’s speculation on my part about where the difficulty sits, and it’s the kind of speculation that gets settled by someone building the thing rather than by argument. But if it’s right, then “AI can’t reproduce taste” is a statement about data collection, not about machines and souls, and it has a much shorter shelf life than people assume.

Where this could still be wrong

Paul Graham’s Taste for Makers makes the opposite case and makes it well: taste isn’t personal preference, and you can tell, because as you get better your old taste doesn’t look merely different to you — it looks worse. Preferences that can be wrong in hindsight aren’t preferences.

I don’t think that contradicts what I’ve said so much as it dates it. “Your old taste was worse” is a verdict delivered later, by a standard that had formed in the meantime. That’s exactly the mechanism I’m describing: no wrong answer at the time, a clear one afterwards. Which raises the real objection to my own distinction. If taste is just judgment with a longer feedback loop, then I’ve described a difference in latency and dressed it up as a difference in kind.

Maybe. Latency isn’t nothing; a skill you can only validate in years behaves differently from one you can validate in a week, in how it’s taught, hired for, and faked.

Two things would settle it. One is a domain where the standard never converges at all, which would either break the distinction or confirm it, and I don’t have that case yet. The other is the evidence this piece is running without: a specific room, a specific call, someone seeing what the rest of us didn’t. I’ve now been writing around that gap for two posts.