Taste is judgment, not preference — and AI can’t fake it

AI can make anything now. A landing page, a dashboard, a brand — ten variants before lunch, each one polished enough to post.

That’s the problem. I’ve written that AI has no taste — and that’s your advantage — and this week made it concrete: when everyone generates, the winners aren’t the ones who generate more. They’re the ones who decide better.

Taste is judgment under constraints — knowing what to leave out, what to protect, and what “good” means for one specific user — and AI can’t fake it because it has no stake in the outcome.

Why does taste matter more now than execution?

Because execution stopped being scarce.

In 2023 a solo founder needed weeks to ship a credible MVP. In 2026 you prompt it before coffee. McKinsey’s 2025 State of AI report found 88% of organizations now use AI in at least one function — the floor rose for everyone at once.

And floors that rise produce convergence. A 2024 study in Science Advances on how generative AI reduces collective diversity found AI helped individuals make more varied ideas alone, but groups using the same models produced work that was more alike, not less. More polished, less distinct.

That’s what you’re seeing in the wild: five new AI products, same blue gradient, same thin icons. The New York Times piece on why taste is the one thing AI can’t replace framed it well — when any idea can be turned into reality, most ideas shouldn’t be. Taste decides which ones deserve to exist.

What does taste actually mean in an AI product?

Taste isn’t “I like this blue.” It’s judgment you can explain. It’s three calls on every surface:

What’s the job? A pricing page’s job is comparison, not admiration. If your agent buries the prices below a hero, it failed the job even if it looks nice. Vercel’s design.md starts with reader and job for exactly this reason — hierarchy follows intent.

What’s left out? Good products say no quickly. No extra card, no decorative gradient, no feature that answers a question no one asked. When I cut a page from 7 blocks to 3, comprehension went up. That’s taste — subtraction with a reason.

What’s the failure mode? AI drafts assume the happy path. Taste asks: what does this look like empty or on the second visit? If there’s no undo or honest copy about limits, the interface is borrowing trust it hasn’t earned.

How do you make taste visible to an agent?

Agents don’t have taste. They have context.

If your taste lives in your head, the model hands you the median. If it lives where the agent can read it, you get closer to what you’d ship. That takes three moves:

Write one page of judgment. Not “be clean and modern.” Write who it’s for, what they came to decide, and what order earns that decision. “Lead with recommendation and both prices on the same scale” is judgment. “Minimal and premium” is not.

Put mechanics in the system. Spacing, type scale, table width — these aren’t opinions to debate each time. They’re classes and variables the agent must use. Figma’s announcement that the canvas is now open to agents matters here: agents that read your real components drift less than agents that invent new ones.

Make proof cheap. If you can test it, don’t ask for it — check it. A grep for a banned pattern, a width assertion, a five-item eval set with known good outputs. Guidance tells the agent what you want. Checks prove it happened.

Vercel ran this loop 200+ times. The point was the split: judgment in a file, mechanics in a stylesheet, proof in code. Copy that split for one artifact and you’ll re-fix the same drift less often.

How do you practice judgment when AI keeps generating?

You slow down where trust is built.

Speed helps for the first 80%. It hurts the last 20%. The founders shipping well now keep a small, strict habit:

Keep the first draft. Don’t re-roll. Save what the model gave you unedited. That’s your baseline — proof of the default.

Write why it’s wrong in three sentences. Hierarchy is flat. Copy promises more than the product does. Empty state has no next step. If you can’t name it, you can’t teach it.

Fix it and save the diff. The diff is the taste file. After ten diffs you’ll see the same three corrections repeating. Those become your one-page guidance and one check. That’s not bureaucracy — that’s memory.

Then gate shipping on judgment, not generation. “You decide” is fine for a first pass, but someone has to hold the gate before it goes live. Ask: does a returning user understand what changed and how to undo it? If not, don’t ship.

Taste looks like preference from the outside. Inside it’s discipline — saying no ten times so the one yes lands.