Toward Computational Taste: LLMs, Aesthetics & Judgment
Toward Computational Taste: LLMs, Aesthetics & Judgment A.Atherton (2025)
Ten years ago, the internet still felt like a place you visited. Taste existed as something humans curated for one another: fashion bloggers posted flat lays, museum curators selected exhibitions, editors filtered and packaged culture into glossy magazines. The digital world was engineered to showcase human taste. Today, the digital world is engineered to learn it.
But taste is slippery. One of the paradoxes of preference is that we crave novelty, and novelty reshapes our preferences. As Oscar Wilde quipped, fashion is “a form of ugliness so absolutely unbearable that we have to alter it every six months.” We chase the new, and in doing so, our idea of what’s good or beautiful shifts. Taste is not fixed, it’s dynamic, adaptive, and socially constructed.
This is what makes computational taste so fascinating: it forces us to ask whether machines can model something that is designed to change.
From People of Taste to Machines of Taste
For centuries, taste was framed as an individual faculty, the “ability to discern what is of good quality or of a high aesthetic standard.” Immanuel Kant treated taste as a universal sense of judgment. Pierre Bourdieu reframed it as social: a reflection of habitus (the dispositions shaped by upbringing and education) and a tool of distinction, used to signal belonging to a class or culture.
Tom Vanderbilt, in You May Also Like, describes taste as operating like traffic: a noisy feedback chamber where one does what others do and vice versa, impossible to predict in the particulars, yet reliably producing flows, just as a certain number of cars travel a road each day, a certain number of songs rise to the Hot 100.
In the analog world, taste was mediated by gatekeepers. In the internet age, taste became participatory. If mass media created audiences, social media let audiences create ever more audiences, Ortega called this “the increase of life.” The internet supercharged social learning, Gabriel Tarde’s term for our instinct to imitate others. Culture became exponential.
Now, large language models and generative systems represent a third regime: not people curating taste for others, nor audiences amplifying each other, but machines learning and reinforcing taste themselves.
This is a profound shift. It means taste is no longer just something people express, it’s something models optimize for.
Why Taste Matters for LLMs
Personalizing large language models (LLMs) to accommodate diverse user preferences is becoming essential. Traditional reinforcement learning from human feedback (RLHF) assumes a single “human” reward model, but that collapses the richness of taste into a monolith.
Recent work like LoRe (Low-Rank Reward Modeling) is showing how to model individual preferences as combinations of shared basis functions, enabling LLMs to adapt to users’ unique aesthetic and stylistic tendencies without retraining. Datasets like Pick-a-Pic, ImageReward, and HPSv3 are doing something similar for visual models: capturing human judgments of style, quality, and appeal to train reward models that approximate taste.
Taste is hard to measure, but it’s no longer invisible. It can now be encoded as data, modeled as a function, optimized as a reward.
This opens the door to an entirely new class of products: LLMs trained not just to be helpful, but to have taste.
Here are five areas where I see opportunities for consumer platforms to emerge at the application layer:
The Meaning of Taste: Aesthetic in the age of AI A. Atherton (2025)
1) Taste Engines: Personalized Judgment as a Service
If today’s recommender systems are blunt instruments, tomorrow’s taste engines will be surgical. Instead of surfacing content based on popularity or similarity, they will curate based on judgment, context-aware, style-specific, and deeply personalized.
Imagine an LLM that builds a multidimensional taste profile for each user across aesthetics, tone, mood, cultural references, and social context. It could rank not just which film you might watch, but which you’d want to quote; not just which product you’d buy, but which aligns with your aspirational self.
Technically, this requires reward modeling from implicit signals (what you dwell on) combined with explicit preference data. LoRe and collaborative ranking approaches make this feasible: they can generalize from a few samples to new domains.
The business opportunity: a Taste-as-a-Service API that powers personalized feeds, recommendations, and rankings across industries shopping, media, dating, travel, even hiring. Think Stripe for taste. Every platform that currently uses collaborative filtering could plug in a taste model and get 10x better engagement.
2) Aesthetic LLMs: Style-Conditioned Generation
Most LLMs today are judged on truthfulness, coherence, and helpfulness but not tastefulness. Yet in domains like marketing, writing, design, and entertainment, taste is the currency.
Models like TAPO (Textual Aesthetics Preference Optimization) and G-Eval are pioneering ways to train LLMs on human-labeled “aesthetic” data. The result is writing that flows better, feels more polished, and aligns with human expectations of beauty.
This unlocks an opportunity for style-conditioned generation: models that can mimic not just the content of what someone might say, but the judgment embedded in how they say it. Imagine choosing from a library of taste profiles, minimalist, maximalist, romantic, avant-garde, and having your content automatically rewritten in that style.
A company here could offer creative teams a collaborative AI “taste partner” that enforces brand aesthetics, tone, and sensibilities across every output, from emails to product packaging. Think of it as Figma for taste.
3) Taste Tribes: Social Networks Sorted by Style
If Bourdieu is right that taste expresses our vision of the “good life,” then taste is inherently social. People cluster around aesthetic worldviews cottagecore, brutalist, streetwear, quiet luxury which serve as signals of identity and belonging.
Today’s social platforms flatten these taste signals into generic engagement metrics. Tomorrow’s platforms will foreground taste as the organizing principle.
Picture a social network where your feed, friends, and groups are organized not by who you know, but by your evolving taste fingerprint. The system could infer your taste profile from your behavior and continually remix your “taste tribe” as your preferences shift.
This creates fertile ground for new consumer platforms: fashion discovery networks, visual bookmarking sites, even dating apps where taste alignment is the primary match signal. Taste could become the new graph.
4) The Taste Graph: Cultural Capital as Data
For centuries, taste has been a proxy for cultural capital, a way to signal education, class, and refinement. Bourdieu argued that aesthetic judgment isn’t neutral: it reflects the values of the groups we aspire to.
As more of culture becomes digitized and labeled, taste can now be mapped as structured data. Every follow, like, purchase, and click becomes a signal of cultural capital.
This creates space for a Taste Graph, a structured ontology of cultural objects, styles, and references connected by who likes them. Companies could use it to build “taste scores” for users, creators, and brands, similar to how PageRank revolutionized search.
This would be invaluable for advertisers, recommendation engines, and creators but also risky, because it makes cultural capital computable. Whoever builds the taste graph could shape culture itself.
5) Taste Alignment: The Personalization Frontier
Finally, the holy grail: aligning models to individual taste at scale.
Mostcurrent RLHF pipelines assume a single, universal definition of “good.” But taste is not universal. It’s personal, plural, and contextual. What’s tasteful to one user is bland to another.
Emerging methods like LoRe show how to solve this: by modeling individual preferences as combinations of shared basis functions, you can personalize LLMs to each user’s taste with just a handful of examples. This sidesteps the need for full fine-tuning and avoids hardcoding users into static categories.
The business model here could look like an LLM taste layer, a personalization API that sits on top of foundation models and adapts them to each user’s preferences. Think of it as a personal aesthetic OS: every interaction with the model refines its sense of your taste, which then informs all future generations.
This could power everything from personal shopping agents to dating concierges to creative copilots in any domain where taste matters more than correctness.
The Stakes of Taste
Taste has always been both deeply personal and profoundly social. It emerges from a mix of individual experience, social imitation, and cultural signaling. It evolves as we do.
Timothy Wilson’s research shows that people underestimate how much they will change; they think the present is the “finished” version of themselves, even though their preferences will keep shifting. This malleability is adaptive: it allows us to explore, mate, survive, and thrive.
That’s what makes the project of computational taste so audacious. It asks machines to model something fluid, performative, and recursive: our sense of what is good, beautiful, and worth aspiring to while that sense is constantly changing.
The risk is that in trying to predict taste, we end up freezing it, locking people into their past selves, reinforcing what they’ve liked rather than helping them discover what they could like. The opportunity is to design systems that do the opposite: that expose people to the unfamiliar, that evolve with their users, and that nudge them toward richer and more expansive aesthetic worlds.
A New Regime of Taste
The internet killed analog taste. Social media fragmented mass taste. Now, LLMs will generate taste.
Just as capitalism, as Joseph Schumpeter wrote, exists to “teach people to want new things,” this next wave of AI systems will not just reflect our preferences, they will shape them. And as they do, they will redefine cultural capital, aesthetic judgment, and even our sense of identity.
We are entering a new regime: one where taste is not just expressed by people, but engineered by machines.
This is a massive cultural and economic shift and a massive opportunity. If you are building toward this future, creating systems that can perceive, model, and evolve with human taste, please get in touch.