TL;DR:
- Generation is no longer scarce. Everypixel estimated that more than 15 billion images were produced by AI tools in under two years, at an average of roughly 34 million a day. When any competent output is a prompt away, the bottleneck moves from making to choosing.
- The scarce skill is taste: the ability to tell good from great, to choose the one right option out of a thousand plausible ones, and to defend the choice in specific terms. This article treats taste as a stack of five trainable layers, not a gift you either have or do not.
- There is a real risk hiding inside the abundance. A 2024 study in Science Advances found that generative AI made individual writing more creative but also made the resulting stories more similar to one another, reducing collective diversity. Fluency without judgment converges toward the average.
- Taste is trainable, but not by exposure alone. Decades of deliberate-practice research show the defining ingredient is structured feedback, not repetition. Meta-analyses put the share of expert performance explained by deliberate practice at roughly 30 to 34 percent – large, but only when the practice includes real evaluation.
- Run the stack like a CEO and a student at the same time. Own the final decision the way a CEO owns a call, and keep upgrading your eye the way a student assumes the current version is only a draft.
For most of history, the hard part of creative and knowledge work was production. Writing the copy, designing the layout, drafting the analysis, shooting the photo. Skill meant being able to make the thing at all, and the people who could make it well were rare enough to be valuable by default.
That world is ending. In early 2024, McKinsey reported that 65 percent of organizations were regularly using generative AI, close to double the figure from its survey just ten months earlier. When two out of three organizations can generate a competent draft on demand, competent drafts stop being a differentiator. Everyone has the factory now. What almost no one has is a reliable way to look at the factory’s output and know which piece is worth keeping.
That skill has an old name: taste. And the central claim of this piece is that taste is not a personality trait you are born with or without. It is a stack of specific, trainable capabilities. You can climb it deliberately, and in the AI era climbing it is one of the highest-return things you can do with your time.
Why abundance makes judgment the moat
There is a comfortable assumption that when tools get better, the human job just moves up a level and everything is fine. The data complicates that story.
The Doshi and Hauser study published in Science Advances in 2024 is the sharpest illustration. The researchers ran an experiment in which some writers were given story ideas from a large language model and others were not. The AI-assisted stories were rated as more creative, better written and more enjoyable, and the effect was strongest for writers who started out less creative. That is the optimistic headline. The uncomfortable part is what happened to the group as a whole: the AI-assisted stories were more similar to each other than the human-only stories. Individual quality went up while collective diversity went down.
This is exactly what you would expect from a tool trained to predict the most probable next token. It pulls everyone toward the same competent middle. The output is fluent, safe and forgettable, and because it is fluent it is easy to mistake for good.
That is the strategic opening. When generation is free and converges on the average, the two things that stay scarce are the ability to recognize what is above average and the willingness to choose it over what is merely fine. Both are functions of taste. This is the CEO-and-student thesis in its purest form: the CEO half is owning the decision about what is good enough to ship, and the student half is continuously sharpening the eye that makes the decision. Neither half is automatable, because the machine has no stake in the outcome and no point of view to defend.
What the research says (verified public data)
| Finding | Source | What it means for taste |
|---|---|---|
| More than 15 billion AI images generated in under two years, averaging about 34 million per day | Everypixel image statistics, 2023 | Output is no longer scarce; the bottleneck moves from producing to selecting |
| AI-assisted stories rated more creative individually, yet more similar to each other, reducing collective diversity | Doshi and Hauser, Science Advances, 2024 | Fluency converges on the average; a defensible point of view becomes the differentiator |
| 65 percent of organizations regularly use generative AI, roughly double in ten months | McKinsey, State of AI, early 2024 | Generation is now a default capability everyone shares; judgment is not |
| Deliberate practice explains roughly 30 to 34 percent of expert performance, and its defining feature is feedback rather than repetition | Ericsson et al. 1993; Macnamara and colleagues, 2014 | Taste is trainable, but only through structured feedback loops, not exposure alone |
The Taste Stack: five layers you can climb
The mistake most people make with taste is treating it as a single thing you either have or lack. It is more useful to see it as a stack, where each layer depends on the one below it. You cannot articulate why something works if you cannot first perceive the difference, and you cannot direct a coherent body of work if you cannot make individual judgment calls.
The table below is an original CEOtudent framework, not a dataset. It breaks taste into five layers, from the foundation up, and pairs each with the CEO decision it forces you to own and the student practice that builds it. Read it from the bottom, because that is the order in which the skill actually develops.
The Taste Stack (CEOtudent editorial framework)
| Layer | What it is | The CEO decision (own it) | The student practice (build it) |
|---|---|---|---|
| 5. Direction | Imposing one coherent point of view across a whole body of work, and editing everything that does not serve it | Decide what your work stands for, and cut the 90 percent that dilutes it | Study bodies of work, not single pieces; ask what the through-line is and whether yours has one |
| 4. Judgment | Choosing the one right option out of many plausible ones, under real constraints | Make the call, ship it, and own the result instead of hedging | Decide on a deadline; afterward, review which calls held up and which did not |
| 3. Articulation | Explaining in specific terms why one option is better than another | Set the criteria the team will be held to, in words, not vibes | Write short critiques; convert a gut reaction into three concrete reasons |
| 2. Discrimination | Perceiving the fine differences between good and great | Refuse to accept fluent-but-generic just because it is fast | Run comparison drills: rank options and name exactly what separates first from second |
| 1. Exposure | The reference library in your head, built from studying the best work in a field | Curate your inputs deliberately instead of consuming whatever the feed serves | Study the top 1 percent on purpose; the average is now infinite and worthless as a teacher |
Notice how the failure modes stack too. Someone stuck at layer one has seen a lot but cannot tell you why anything is good. Someone at layer two can feel that something is off but cannot name it, which means they cannot direct anyone else. The people who are genuinely dangerous in the AI era live at layers four and five: they can look at a hundred AI outputs, kill ninety-five without hesitation, and shape the survivors into something with a spine.
How each layer actually gets built
Layer 1, Exposure. Taste starts as a reference library. You cannot recognize great work if your internal comparison set is made of average work. The AI era makes this trickier, not easier, because the volume of competent, forgettable content is now effectively infinite, and it is the default thing the feed shows you. The deliberate move is to study the top of a field on purpose. Read the essays that other essays cite. Take apart the products people copy. Consuming the average teaches you the average.
Layer 2, Discrimination. This is where perception sharpens from a vague sense into a fine-grained one. The mechanism is comparison. Put two options side by side and force a ranking. Most people never do this; they look at one thing and ask “is this good?” which the brain answers with “it is fine.” Ask instead “which of these two is better, and by exactly how much?” and the eye starts to resolve differences it used to miss. This is the layer AI struggles to reach on your behalf, because a model optimized for the probable middle is not built to prize the exception.
Layer 3, Articulation. Perception you cannot explain is not yet usable, because you cannot teach it, direct with it, or defend it under pressure. Articulation is the discipline of converting a reaction into criteria. The practice is blunt: after you judge something, write down three specific reasons in plain language. Not “it feels clean” but “the hierarchy is clear, there is one focal point, and nothing competes with the headline.” Once your taste lives in words, it can travel to other people and to your own future decisions.
Layer 4, Judgment. This is the layer where taste meets reality. Infinite options are a curse without the willingness to choose one and stand behind it. Judgment is the CEO act inside the stack: you make the call under a deadline and with imperfect information, you ship, and you own what happens. The student half is the review afterward. Deliberate-practice research is unambiguous on this point. What separates people who keep improving from people who plateau is not more repetition, it is structured feedback on whether the call was right. Ship, then look back honestly at which judgments held.
Layer 5, Direction. The top of the stack is a point of view expressed across a whole body of work. Anyone can get one good output from a good prompt. Almost no one can impose a coherent aesthetic across fifty of them and ruthlessly cut everything that does not belong. In a world where the Doshi and Hauser result predicts that most AI-assisted work will drift toward the same competent center, direction is the thing that keeps yours recognizably yours. It is the difference between a pile of fine pieces and a portfolio that means something.
The CEO-and-student way to run the stack
Taste tempts two failure modes, and the CEOtudent frame is designed to avoid both.
The first failure is pure CEO: strong opinions, fast decisions, no learning. This person has taste, or had it once, and now defends every call as a matter of authority. In a field that is being reshaped monthly by new tools, a frozen eye goes stale fast. The second failure is pure student: endless studying, comparison and refinement, but never shipping, never making the call, never getting the feedback that only comes from a decision meeting reality.
The stack works when you run both halves at once. Own each judgment like a CEO, on a deadline, with your name on it. Then interrogate it like a student who assumes the current version of your taste is a draft that the next hundred decisions will revise. That loop – decide, ship, review, adjust – is the same structured-feedback engine that the deliberate-practice literature identifies as the real driver of expertise. The tools will keep getting better at generating. Your job is to keep getting better at choosing.
Frequently asked questions
Is taste actually trainable, or are some people just born with it?
It is trainable, and the mechanism is well studied. The deliberate-practice literature associated with Ericsson and later meta-analyses by Macnamara and colleagues found that structured practice explains a substantial share of expert performance, on the order of 30 percent or more, with the key ingredient being feedback rather than raw repetition. Natural inclination exists, but the gap between a trained eye and an untrained one is mostly built, not inherited.
Will AI eventually develop taste and make this skill obsolete?
AI is very good at producing what is probable and average, which is the opposite of taste. The Science Advances research showed AI-assisted work converging toward similarity even as individual quality rose. Taste is fundamentally about prizing the exception and defending a point of view, and a model with no stake in the outcome and no perspective to protect has no basis for either. The tool can widen your set of options; choosing among them is still your job.
How is this different from just having an opinion?
An opinion is layer two or three at best – a reaction, maybe explained. The full stack includes judgment under real constraints and direction across a body of work. The test is not whether you can say what you like, but whether you can choose the one right thing out of many under a deadline, defend it in specific terms, and hold a consistent line across everything you ship.
Where should someone with no background start?
At the bottom. Build exposure by studying the best work in the field you care about, not the average work the feed serves you. Then start running comparison drills at layer two: rank two options and name exactly why one wins. Those two habits alone move you further than any amount of consuming more content.
Does this only apply to designers and writers?
No. Taste in the general sense is judgment about quality, and it applies anywhere AI can now produce a plausible draft: strategy, code, analysis, product decisions, hiring. Wherever generation has become cheap, the ability to evaluate the generation is where the remaining human value concentrates.
Sources
- Everypixel Journal, “AI Image Statistics,” on the estimate that more than 15 billion images were generated by AI tools in under two years and an average of roughly 34 million images per day following the launch of DALL-E 2.
- Anil R. Doshi and Oliver P. Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content,” Science Advances (2024), on the finding that AI-assisted stories were rated more creative individually yet were more similar to one another, reducing collective diversity.
- McKinsey and Company, “The state of AI in early 2024,” on the finding that 65 percent of surveyed organizations reported regularly using generative AI in at least one business function, close to double the share reported ten months earlier.
- K. Anders Ericsson, Ralf Th. Krampe and Clemens Tesch-Romer, “The Role of Deliberate Practice in the Acquisition of Expert Performance,” Psychological Review (1993), on deliberate practice and the central role of feedback and expert guidance.
- Brooke N. Macnamara, David Z. Hambrick and Frederick L. Oswald, “Deliberate Practice and Performance in Music, Games, Sports, Education, and Professions: A Meta-Analysis,” Psychological Science (2014), on the finding that deliberate practice explains a substantial but partial share of expert performance.
- Herbert A. Simon and William G. Chase, on the classic finding that expertise is built on recognizing a large stored library of meaningful patterns rather than raw processing speed.
- Rick Rubin, “The Creative Act: A Way of Being,” on the practice of paying disciplined attention to one’s own reactions as the foundation of judgment.
This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.
This post is also available in:
