Good Taste the Only Real Moat Left
AI made competent output cheap. You can generate a landing page in minutes, draft a product memo from one prompt, and polish a pitch deck before anyone has decided what the company actually believes.
So taste became a topic in tech. When anyone can ship something that looks decent, the people who stand out are the ones who can tell what is generic, what is true, and what is worth pushing further.
There is a catch. If you reduce yourself to picking the best of what a model produces, you become a reviewer of a machine's work instead of a builder with something at stake. Taste matters more than ever, and taste alone will leave you exposed. The edge is taste plus context, constraints, and the nerve to build something the average could never have produced.
What taste actually means
Taste here has nothing to do with luxury or status. It is judgment under uncertainty.
Most work that matters comes without clean data. No spreadsheet tells you which sentence will make a customer care, which feature earns a month of engineering, or which design slips from polished to forgettable. You decide anyway.
Taste shows up in three places:
- What you notice
- What you reject
- How well you can say what is wrong
The third one carries the weight. Plenty of people can say "this feels off." Few can say "this fails because it reads like every other SaaS homepage," or "this buries a regulatory limit under marketing language and will confuse the customer." Taste earns its keep when it turns from a vibe into a diagnosis you can act on.
Why AI and LLMs flatten the middle
An LLM compresses patterns. It absorbs enormous amounts of language, design, and interface, then recombines it fast. That speed is the strength and the bias in one.
These models predict plausible output. They are weaker at originating something specific to your situation, so left alone they drift toward the safe center of the distribution.
That pull toward the center is why so much AI work feels familiar:
- Landing pages with different logos and the same skeleton
- Product copy that could describe any app
- Essays with tidy headings and no lived judgment
- Design that looks modern and forgettable
None of this is broken. It is average, produced at scale. Average used to take enough effort to create some separation. Now it is everywhere, and a crowded 7-out-of-10 world is the result.
The new bottleneck is judgment
Before AI, mediocre work usually meant someone ran out of time, money, or skill. Now it usually means someone stopped at the first acceptable draft.
AI drives down the cost of that first draft, so the value sits downstream, in what you do after it.
The scarce move is saying:
- This looks fine and it is too generic
- This sounds impressive and it hides the real trade-off
- This interface is polished and it fights how the user thinks
- This plan is ambitious and the operating constraints sink it
Generation is cheap now. Refusal is what stayed scarce.
AI as a mirror for your own taste
AI also holds up a mirror. Ask it for ten versions of a homepage hero, an onboarding flow, a support email, or a product pitch, and you tend to get the same spread:
- A few weak ones
- A big cluster of acceptable ones
- One or two near what you wanted
Skip "which one do I pick?" Ask why most of them are still wrong. How sharply you can answer that is your taste. If the critique stays vague, the taste is still raw. If it gets specific, your judgment now beats the model's output, and you can drive the tool instead of following it.
Here is the division of labor:
| Layer | AI and LLMs do well | Humans still need to do |
|---|---|---|
| Generation | Produce many plausible variations fast | Decide which direction matters |
| Pattern matching | Recombine common structures and phrasing | Spot what is too generic for this situation |
| Optimization | Improve toward a stated target | Decide whether the target itself is right |
| Scaling | Turn one idea into many assets | Carry the real context, stakes, and consequences |
The model hands you options. Ownership stays with you.
A practical loop for training taste
Taste grows through exposure, critique, and shipping. AI can speed that loop up.
The method:
- Pick one high-leverage artifact from your week: a paragraph, a pricing explanation, a dashboard label, a customer email, a slide.
- Generate 10 to 20 versions.
- For each, write one sentence starting with "fails because..."
- Rewrite the best version under a hard constraint:
- No buzzwords
- One idea per sentence
- Must name a real trade-off
- Must make sense to a first-time user
- Ship it somewhere real and watch what happens.
You are not training AI to choose for you. You are building a sharper vocabulary for saying no. Do it enough and your defaults change: you stop admiring polish, and you spot empty specificity, borrowed tone, and fake confidence faster.
Why taste alone is not enough
The "taste matters" argument has a strong version that boxes you into a small role: the model generates, and you stand at the end of the line picking the best one. Useful, and too small.
Most important work never came from picking. It came from making something under constraint, arguing with reality, collaborators, budgets, deadlines, and the cost of getting it wrong. That friction is where depth comes from.
Reduce your job to curation and you become a discriminator in a machine's loop. The GAN analogy is rough but it lands: the discriminator exists to make the generator better, and once the generator is good enough, the discriminator is not what ships. Taste has real value. Taste with no authorship, no stake, and nothing built is a narrow role, and a fragile one.
What humans still do that models cannot own
A model generates, recombines, and optimizes against a prompt. What it can't own is the part of the work that carries consequence.
Three of those parts:
1. Holding the stake
Real products run under stakes that don't fit in a prompt: trust, regulatory exposure, outage risk, team capacity, brand damage, on-call pain. A model can draft copy for a payments feature. It can't answer for that copy when it hides a regulatory limit and the support queue floods.
2. Working with the genuinely new
New ideas tend to look wrong at first. They don't match the training set, so they read as awkward or off-standard. You can sit with that discomfort and protect something fragile long enough for other people to get it.
3. Choosing direction
The biggest calls are directional, not cosmetic. What problem is worth solving? Which trade-off can you live with? What product do you want your name on? What will you refuse to optimize for? That is authorship, and it happens before any prompt.
Why this matters for builders
This goes past any one market, because the temptation is everywhere: ship competent surface and call it meaningful work.
The tools are open to everyone. A small team or a solo builder can now ship what used to take a whole org. Good news, with a catch.
Teams can now produce work that is polished worldwide and hollow up close. A fintech interface can sound sophisticated and never explain timing, settlement, or what support will actually do. A B2B site can look world-class and say nothing a real buyer would recognize as grounded. A devtool can market beautifully and ignore the on-call load, compliance pressure, and cost limits its users live with. AI makes it easy to sound sophisticated and does nothing to make you specific, and specificity is the whole advantage.
For builders, taste means moving toward real context:
- Write for how people understand the problem, not how SaaS templates talk about it
- Pull domain and operating constraints into the product instead of hiding them behind abstraction
- Design for distracted, low-attention, real conditions, not demo conditions
- Use AI to map the canon fast, then leave it where the context demands
The market has enough competent clones. It needs builders who keep AI's speed without giving up the specifics that make a product worth trusting.
A better way to use AI
Passive selection is the weak way to use AI. Active shaping is the other one:
- Explore the design space faster
- Study the best existing work and learn the canon
- Generate alternatives you would not have reached on your own
- Reject what is generic, dishonest, or context-blind
- Add the constraints the model can't know, then build from there
When AI output looks polished and feels hollow, ask:
What am I adding that the model could not add on its own?
Good answers:
- A real operating constraint
- A user truth you learned the hard way
- A regulatory nuance
- A cultural detail
- A strategic trade-off
- A view you will defend in public
If you can't name the addition, you are still just consuming.
Taste as a side-effect of serious work
The honest conclusion is unglamorous. Taste is a by-product of paying close attention to reality. It grows when you:
- Study strong work
- Generate many options and refuse to fall for the first
- Learn to diagnose why something fails
- Ship where feedback has consequences
- Stay close to the domain instead of floating above it
A model makes the first draft cheap. It does not make your judgment automatic, hand you ownership, or choose what deserves to exist. That is why taste matters more now, and why taste on its own still falls short.
The edge in this era is not better vibes than the model. You use the model to clear away average output faster, then spend your judgment where it counts: direction, specificity, consequence, and the nerve to build something the statistical middle could never reach.