If you’ve spent any time in AI hiring conversations lately, you’ve heard it. Investors want founders who have it. Hiring managers say they can’t define it but know it when they see it. Product teams claim it’s the secret weapon that separates transformative AI products from forgettable ones. The word is taste — and it has become, in the span of a few conference cycles, one of the most overloaded and underexamined terms in the AI talent conversation. This piece is an attempt to trace where the idea came from, what it actually means when used precisely, and why so much of how it’s currently deployed is a kind of flattering noise.
Legitimate Origins
The “taste” metaphor in creative and technical domains has deep roots that predate the AI moment entirely. Steve Jobs used it extensively — his formulation was that technology alone is insufficient, that it must be married to the liberal arts and the humanities to produce work that makes “our hearts sing.” He was drawing on a tradition rooted in aesthetics philosophy: the idea that some people have cultivated an internal model of what good looks like that cannot be fully articulated through rules or checklists.
Paul Graham codified this for the hacker tradition in his 2002 essay Taste for Makers, arguing that good taste is an actual cognitive skill — the ability to recognize elegance, simplicity, and rightness — and that it separates great engineers from average ones. This was a meaningful, specific claim: taste as trained intuitive judgment, not mere subjective preference.
Migration Into AI Discourse
The metaphor found new life in the generative AI era for a specific, defensible reason: when tools become powerful enough that what you ask for matters more than how you build it, the discriminating faculty becomes the scarce resource. Andrej Karpathy popularized this framing in the LLM context — the idea that the bottleneck in AI product development is no longer engineering execution but judgment about what’s worth building and what output is genuinely good.
There is something real here. When anyone can invoke a capable model, the person who can tell the difference between a mediocre and an excellent output — and iterate toward the excellent one — has a genuine edge. That is a defensible use of the metaphor, and it explains why the word gained traction in serious technical conversations.
When anyone can invoke a capable model, the person who can distinguish mediocre from excellent — and close that gap — has a genuine edge. That is the defensible core. Everything else is performance.
Where It Becomes a Fad
The problem is that “taste” has become a status signal more than a descriptive term. It is now used interchangeably to mean roughly five different things: product intuition, aesthetic sensibility, domain expertise, curatorial judgment, and vague senior-ness. Hiring managers who say they want someone with taste often mean something closer to “someone whose outputs I won’t be embarrassed by” — which is simply competence, and has a more honest name.
The metaphor also carries implicit class and cultural baggage that rarely gets examined. “Taste” has historically been used to naturalize what are actually learned, often socioeconomically contingent skills. Pierre Bourdieu’s entire sociological project was essentially a deconstruction of taste as a social sorting mechanism disguised as an innate faculty. When a venture capitalist announces they want founders with taste, there is frequently an unexamined cultural fit judgment lurking underneath the aesthetic language — one that can quietly exclude people whose formation was different, not deficient.
The Bourdieu Warning
“Taste” has a sociological history as a mechanism of class reproduction — naturalizing learned, culturally contingent skills as innate faculties. AI hiring that prizes taste without defining it risks importing exactly this bias into an industry that already struggles with homogeneity.
The Useful Core
Strip away the status signaling and something worth preserving remains. The ability to recognize quality before you can fully articulate why something is good — to have internalized enough examples of excellent work that your intuition becomes a reliable compass — is a real and trainable skill. It just isn’t mystical, and it isn’t taste in the sommelier sense.
It is more precisely described as calibration or domain fluency. Have you consumed enough great work in your domain that you carry an accurate internal benchmark? That is teachable, measurable, observable over time, and it does not require inherited cultural capital or the right aesthetic pedigree.
People using the term most precisely tend to mean exactly this: a well-calibrated internal model of quality built through immersion and iteration. People using it most loosely tend to mean: does this person remind me of others I already respect? The former is a hiring criterion. The latter is a bias.
The term “taste” conflates at least five distinct competencies: product intuition, aesthetic sensibility, domain expertise, curatorial judgment, and vague seniority. Only the first four are real, observable skills. The fifth is cultural fit in disguise — and conflating them all under one word makes it impossible to assess or develop any of them deliberately.
What to Do With This
The next time someone tells you they’re looking for AI talent with taste, it’s worth pressing on what they mean. If the answer is something like “they’ve shipped things, they’ve made judgment calls under uncertainty, and their outputs consistently exceed expectations” — that’s a real criterion, and taste is a reasonable shorthand. If the answer circles back to cultural fit, aesthetic alignment with the founder, or the feeling that someone just gets it — the word is doing social work, not evaluative work.
The AI industry would be well served by retiring the term from hiring criteria altogether and replacing it with the specific, observable behaviors that underlie the concept. Calibration. Domain fluency. Evidence of iterating toward quality. These things can be assessed, developed, and taught. Taste, in its current usage, mostly can’t — because it was never really defined in the first place.
