The taste moat is cracking.

On February 14, 2026, Paul Graham posted a prediction to his 2.2 million followers: “In the AI age, taste will become even more important. When anyone can make anything, the big differentiator is what you choose to make.” He linked his 2002 essay Taste For Makers. Two days later, OpenAI president Greg Brockman replied with five words: “Taste is a new core skill.” The exchange reached 3.7 million impressions, and by spring, “taste is the moat” had hardened into the default answer to what survives AI.

Six months later the line is everywhere: pitch decks, hiring posts, every founder thread on X. That ubiquity should worry you. A moat 3.7 million people have been told to dig is not a moat, it is a lane. And the sharpest pushback is not coming from skeptics outside the industry. It is coming from the people shipping the models.

The counter-quote comes from inside the lab

Jenny Wen leads design for Claude at Anthropic; before that she was a director of design at Figma. She is no bystander in this debate: her team ships one of the tools the taste-is-the-moat crowd is bracing against. Her warning to her own profession is blunt. Designers, she says, may be holding onto taste and judgment as a moat “a little too tightly,” because AI is going to keep getting better at exactly that.

Her account of her own calendar backs it up. “This design process that designers have been taught, we sort of treat it as gospel. That’s basically dead,” she said in a March interview. A few years ago, mocking and prototyping took 60 to 70% of her time. Now it takes 30 to 40%, and the recovered hours go to pairing directly with engineers. Read that shift carefully. The deliverable that most visibly embodied a designer’s taste, the polished mock, is what the tools ate first. What survived is the part where judgment enters someone else’s system.

Taste as a vibe is already commoditized

If you want to see what happens to taste once it becomes legible, look at what cultural theorist Emily Segal named “tasteslop” in May 2026: AI-generated brand moodboards recombining the same fetish objects of modernist taste. A Rimowa suitcase. A skinny-neck kettle. A Dieter Rams monograph. Her one-line history of how we got here: “Instagram made taste visible, algorithms made it repeatable, and AI made it scalable.”

Segal’s decomposition is the useful part. She splits good taste into discernment (an articulated judgment about why one thing beats another), pattern recognition (historical knowledge plus intuition), and idiosyncrasy (the personal, contextual part that resists cloning). Notice which two of the three are learnable from data. Pattern recognition is what a model is. Discernment, once written down, is training data. The only component that resists scale is idiosyncrasy, and idiosyncrasy on its own is a personal brand, not a defensible business.

The numbers describe a split, not a slaughter

Figma’s State of the Designer 2026 report, run with research partner NewtonX, surveyed 906 designers across North America, APAC, Europe, LATAM, and the Middle East. The productivity numbers are lopsided: 91% say AI tools improve their designs, 89% report working faster. Then the consensus ends. Asked whether the profession has gotten better or worse, 36% say better, 35% say worse, and the rest say about the same.

Both camps are reporting accurately on their own position. When 91% of practitioners get better output from the same tools, the baseline rises for everyone at once, and taste-as-baseline stops differentiating anyone. The designers reporting “better” look, on the evidence of Wen’s calendar, like the ones whose judgment moved upstream: into systems, reviews, and pairing. The ones reporting “worse” are competing on the deliverable the tools now produce by default.

The moat that holds is a system

Here is the reframe, and engineers are better positioned for it than designers. Personal taste is a vibe. It scales to one person’s working hours, walks out of the building when they do, and depreciates as models close the gap. Judgment at scale is infrastructure: written down, versioned, enforced by tooling, running on every generation whether you are in the room or not. Wen’s time shift is the pattern in miniature. Less time producing tasteful artifacts by hand, more time installing judgment where the production actually happens.

Concretely, judgment at scale looks like this:

• Written rubrics, not gut calls. If you can articulate why one interface beats another, that judgment can gate a pipeline instead of living in your head. Our taste rubric is a working example: criteria specific enough to score against.
• Exemplars in the repo. Models follow reference implementations far more faithfully than adjectives. Three annotated screens you consider excellent will steer a coding agent better than a paragraph of brand values.
• Design decisions encoded where tools read them. Type ramps, spacing scales, color tokens, and forbidden patterns written into the files your agents load. A decision made once and enforced ten thousand times outcompetes a great eye applied ad hoc.
• Review loops that look at pixels. Screenshot checks and design tests that fail loudly. Taste that never inspects the rendered output is opinion, not quality control.

None of this makes the eye obsolete. Someone still has to author the rubric, choose the exemplars, and notice what the checks miss, and the sharper that person’s discernment, the better the whole system compounds. That is why we teach the eye first: the Deconstructing Great Interfaces series trains you to name what makes an interface work, and the design taste resources page maps the wider curriculum. The difference is what you do with the discernment once you have it. Kept as intuition, it is a wasting asset. Written into a system, it appreciates with every model release, because better models execute your encoded judgment more faithfully.

Graham was half right

The February prediction holds up better than the slogan it spawned. Taste does matter more when anyone can make anything; choosing what to make is a genuine differentiator. What does not hold up is the comforting corollary everyone attached to it: that having taste is a defensible position. Wen’s point lands here. The models are being aimed at judgment itself, by teams that include people like her, and “I have a good eye” is precisely the kind of pattern a model learns to imitate. The founders still repeating the February line are defending a vibe. The engineers who come out ahead over the next two years will be the ones who turned their judgment into a system their tools can run, then kept sharpening the judgment the system encodes. Build the eye, then build the machine that ships it.