Why the AI era will reward discernment, not merely production
“I know what I like and what I don’t like. And I’m decisive about what I like and what I don’t like.” Rick Rubin, speaking to 60 Minutes
Rick Rubin barely plays an instrument. He does not operate a soundboard. Yet, for four decades, artists have invited him into the room when the work matters.
The reason is not mystical. Rubin is unusually good at deciding. He can hear a half-finished track, notice the accidental riff between two demos, and tell an artist what to cut. He calls himself a reducer rather than a producer. In a culture now obsessed with generation, that is a useful job description.
Generative AI is about to make the distinction unavoidable. The tools can produce text, images, code, strategy decks and confident first drafts at a speed that would have sounded implausible a few years ago. The interesting question is no longer whether they can make things.
It is whether the person using them knows what to keep, what to reject and what they are trying to say before the machine starts talking.
Rick Rubin’s job is not vibes
There is an unhelpful way to use Rubin as an example. Call taste ineffable. Put a candle beside a moodboard. Tell the client that some people simply have it.
His career points somewhere more practical.
Taste is not decorative preference. It is informed discrimination. It is recognising the difference between what is technically impressive and what is emotionally honest. It is built through exposure, comparison, failed attempts, context and the awkward habit of being specific when everyone else is being politely vague.
Rubin’s story about Tom Petty shows the point. He heard a guitar riff played between demo songs and thought it was the best thing on the tape. Petty turned it into “Last Dance with Mary Jane.” That was attention, backed by listening and a willingness to make a judgment that could have been wrong.[1]
The analogy matters for AI. A model can offer fifty possible routes. It cannot carry the same responsibility for choosing one route, in one cultural moment, for one audience, with one set of consequences.
A prompt can be a creative act. It is not the whole act. Before the prompt comes a more difficult question: what do I want to make, and why this rather than the forty-nine plausible alternatives?
Cheap execution moves the bottleneck
One of the supplied videos quotes an investor proposition that is worth taking seriously: “when intelligence gets cheaper, taste becomes more valuable.”[2]
It is a neat line. It needs a little care. Intelligence is not simply becoming cheap. What is becoming cheaper, faster and easier to access is much of the first pass: production, analysis, ideation and iteration. Stanford’s 2025 AI Index reports that the inference cost of a system performing at GPT-3.5 level fell by more than 280-fold between November 2022 and October 2024.[3]
That does not make judgment evenly distributed. It changes the place where value sits.
| When execution was scarce | When execution is abundant |
|---|---|
| The advantage was access to production capacity | The advantage shifts toward framing and selection |
| A technically clean first draft was valuable | A technically clean first draft is a starting point |
| Specialists controlled tools and workflows | More people can create plausible outputs |
| The bottleneck was making | The bottleneck is deciding what deserves attention |
This matters to brands as much as it matters to artists. If every competitor can generate a campaign territory before lunch, a polished deck stops proving strategic depth. If every product team can create twenty interface directions, the harder task is explaining why eighteen of them are wrong for the people who will use them.
AI expands the option set. Taste supplies the criterion.
Without a criterion, more becomes the default strategy: more content, more variations, more dashboards, more persona documents and more things nobody had the nerve to exclude. The machine can flood the room. It cannot decide which conversation is worth having.
The similarity tax nobody puts in the forecast
There is a tempting story about AI and creativity: give everyone a brilliant assistant, and everyone becomes more original. The evidence is less tidy.
A preregistered experiment published in Science Advances asked 293 people to write short stories, either without AI or with one or up to five ideas from GPT-4. Six hundred evaluators then assessed the results. Stories created with AI ideas were rated as more creative, better written and more enjoyable, particularly for people who started out with lower creativity scores. The AI-assisted stories were also more similar to one another.[4]
That is the trade-off worth keeping in view. We may get a better average output and a narrower cultural range.
This is not an argument for banning the tools or pretending that a blank page makes us morally superior. A writer, designer or strategist should use a good assistant when it helps them explore, test and refine. But individual improvement can coexist with collective sameness.
The risk is not that every AI output looks bad. Most of the time, the risk is more subtle. So much of it looks acceptable that we lose sensitivity to the difference between acceptable and necessary.
Taste is a defence against that convergence. It asks questions that slow the room down: Have I seen this before? Is it right for this subject, or merely close to the average of what the training data has seen? Is it useful here, or simply fluent? What disappeared because the output arrived too quickly?
David Droga puts the point more generously. In his conversation with Wharton, he argues that AI can help creative teams reach bad ideas faster and move on from them. Good. Someone still has to recognise a bad idea when it arrives wearing a very convincing suit.[5]
Taste has to be learnt
The word “taste” can sound elitist, as though it belongs to people with expensive record collections who speak in low light. In practice, it is learned.
It grows through work. You compare versions, see your first instinct fail, receive feedback and build a reference library larger than whatever the model happened to predict next. That is why this conversation connects directly to the apprentice problem.
If AI absorbs too much early, repetitive work, it may also erase some of the situations in which people acquire the patterns that later become taste. A junior creative needs to write bad headlines before they can hear why a good one has tension. A young strategist needs to sit through weak research before they can feel the absence of a relevant question. A product person needs to watch somebody misunderstand an interface before they can tell the difference between elegant and usable.
Research into human-AI co-creation makes a similar distinction. AI is useful for divergent exploration, but designers still have to filter, adapt, recombine and set the final direction. More experienced designers use AI selectively for visualisation and refinement. Less experienced designers use it more often to help frame the problem in the first place.[6]
That has consequences for how companies adopt these tools.
| Weak AI practice | Taste-building AI practice |
|---|---|
| Ask the model for the answer | Ask people to state a hypothesis before consulting the model |
| Generate one output and polish it | Generate alternatives, then explain the criteria for rejecting them |
| Treat review as a formatting check | Review assumptions, omissions and cultural fit |
| Reward volume and speed | Reward clear choices and accountable rationale |
The important thing is not an “AI approval gate” added to every meeting. It is making selection visible. Senior people need to explain why an apparently competent answer fails the brief. Teams need to retain rejected options long enough to learn from them. The reasoning cannot stay private theatre in the head of the most experienced person in the room.
Choosing is a moral act
Taste has a quieter ethical dimension too.
When an organisation lets a model decide what appears first, which customer story is typical, which applicant looks relevant or which cultural reference is safe, it is still making a judgment. It has simply handed the judgment to patterns shaped by historical data, commercial incentives and defaults that most people in the room cannot see.
The useful human intervention is not a person placed at the end of a pipeline so the company can say “human in the loop.” The work starts earlier. It means making the values behind selection explicit before the pipeline begins.
Whose voice does this output make easier to hear? Whose experience becomes less legible? What does good mean here? What should the brand refuse to optimise, even if the model can scale it?
Those are taste questions. They are also leadership questions.
Rick Rubin’s lesson is not that intuition beats craft. It is that craft becomes more powerful when someone protects the signal from the noise. In the AI era, the capacity to do that may be more valuable than the capacity to make another perfectly adequate thing.
The companies that matter will not be the ones that generate the most. They will be the ones that can say what they will leave out, what they will pursue and what they are willing to be known for.
Cui bono when every creative decision is delegated to a system designed to predict what has already worked?
References
[1] CBS News, In Shangri-La with music producer Rick Rubin, 28 May 2023.
[2] Twill Socials SF, Instagram Reel, citing a 20VC conversation with Geoffrey Woo, Jake Paul and Harry Stebbings. The line is an investor thesis, not empirical proof.
[3] Stanford Institute for Human-Centered Artificial Intelligence, AI Index Report 2025, 2025.
[4] Anil R. Doshi and Oliver P. Hauser, Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances, 2024.
[5] Wharton AI & Analytics, AI, Taste, and the Future of Creativity: David Droga on What’s Worth Saving, 22 July 2025.
[6] N. Wang et al., Exploring creativity in human-AI co-creation, Frontiers in Computer Science, 2025.