Taste is not a hidden numerical preference waiting for sufficient data. It is cultivated judgement — historical, embodied, socially answerable — and models may extend it without owning it.
Recommendation culture has a flattering ontology of the self: you are a vector, slightly noisy, increasingly knowable. Taste, in that ontology, is a setting — a latent ranking over items, recoverable if the system watches long enough. The flattery is that your uniqueness can be optimised for. The cost is that judgement becomes a preference curve, and preference curves do not have to answer for themselves in public.
The institute rejects the setting metaphor without rejecting recommendation as engineering. Filters can be useful. Predictions can be useful. What fails is the philosophical export: the claim that what people mean by taste is what collaborative filtering means by preference. Knowledge-how literature already marks a nearby boundary — practical knowing is not always reducible to piles of propositions. Taste sits closer to cultivated how than to stored that.
Source ·Knowledge How (opens in new tab)· Stanford Encyclopedia of Philosophyatlas
1 — Cultivation, not decryption
Taste is learned in time. It is shaped by encounters, apprenticeships, refusals, accidents of geography, and the slow education of attention. It is historical: what counts as elegant, rigorous, funny, or cheap is indexed to traditions and counter-traditions. It is embodied: the flinch at a typeface, the patience for a long proof, the intolerance for a certain kind of slide, are not always available as declarative reasons at the moment they operate. It is socially answerable: to claim taste is often to claim a standing from which one can be challenged — by a peer, a public, a craft.
A setting does not get challenged. It gets updated. That single grammatical difference should make product teams nervous whenever they reach for "taste" as a feature name. Updating is private and painless. Challenge is public and sometimes humiliating. Taste without the possibility of humiliation is cosplay.
The comic product translation is a slider labelled "more tasteful." The slider always moves the output toward the vendor's prior while congratulating the user for self-knowledge. This is not empowerment. It is autocomplete with a mirror attached.
2 — What models can extend
None of this implies that machine systems cannot participate in taste-laden work. They can draft in a register, propose alternatives, surface references, violate a local style on purpose, or help a human notice an inconsistency in their own standards. Extension is real. Ownership is not transferred by extension.
A model does not bear the social answerability that taste, in the full sense, includes. It can be fine-tuned toward a corpus that embeds a community's judgements. That is inheritance of sediment, not possession of standing. When a system says "this is better," the useful reading is usually "this matches a pattern that was rewarded." When a person says "this is better" inside a craft, they can be asked why, and the asking is part of how the craft continues.
Confusing sediment with standing is how organisations end up treating model rankings as critical authority. The institute prefers a colder description: high-probability continuation under a style constraint. Useful. Not a curator. The colder description also protects the model from a kind of blame it cannot meaningfully carry. If a ranking is sediment, the accountability stays with whoever chose the corpus, the objective, and the deployment context. Animating the model as a taster is a way of misplacing the signature line.
3 — Against taste as personalisation destiny
Personalisation promises to fit the world to you. Taste, as cultivation, often works the other way: you are remade by standards you did not invent. Education is not only preference satisfaction. Criticism is not only matching. Sometimes the point of a tradition is to make certain preferences harder to hold without embarrassment.
Another worry: if taste is not a setting, platforms cannot productise it cleanly. Correct. Some human goods resist clean productisation. The failure mode is not that engineering stops. The failure mode is that engineering redefines the good until it fits the database, then sells the database as self-knowledge.
4 — Implications for interfaces and institutes
Design consequences follow without a manifesto.
Do not present preference capture as the recovery of taste. Say what you are doing: predicting clicks, ranking engagement, matching embeddings. Do not hide collective priors behind second-person address ("because you love…") when the prior is doing the labour. Offer tools that support cultivation: comparison, history, slow looking, exposure to dispreferred exemplars, structured disagreement — not only more of the same basin.
For AI assistance in creative and intellectual work, prefer interfaces that keep the human on the hook for evaluative claims. The model may propose; the user must be able to own. Where organisations use models to enforce a house style, call it house style enforcement. That is at least honest. Calling it "aligning to taste" smuggles a moral psychology the system does not have.
There is a research-adjacent temptation to treat revealed preferences at scale as the democratisation of aesthetics. Scale can reveal regularities. Regularities are not yet justifications. A majority cluster in embedding space can tell you what is common; it cannot tell you what is worth defending when commonness is the problem. Taste, when it is doing its hardest work, is often anti-average on purpose — and "anti-average" is not a parameter you set once.
The related notes on generative monoculture and on logic as a local maximum belong in the same drawer. Homogenization is what happens when cultivated difference is expensive and high-probability sameness is cheap. Local maxima are what happen when a frame is mistaken for the world. Taste is one of the human practices that can notice both — imperfectly, argumentatively, without a guarantee. Replacing it with a setting does not democratise judgement. It deletes the scene in which judgement could be answerable.
“The authors deny that taste is a setting, then ask us to prefer their settings. Reject unless they ship the slider they mock.”
We will not ship the slider. We will continue to argue in public, which is a worse business model and a better epistemology. Taste remains a practice you can be wrong about in front of other people. That risk is not a bug in the human stack. It is one of the ways the stack stays alive — and remains answerable to others.