As AI systems grow more fluent and personalised, researchers are confronting a deeper challenge: understanding human taste itself, not merely predicting preferences, but modeling culture, identity, memory, and meaning computationally
Modern large language models can summarise research papers, generate poetry, and simulate human conversation with remarkable fluency. Yet despite this sophistication, much of AI-generated content still feels strangely generic. The prose is coherent but forgettable. The recommendations are relevant but emotionally flat.
The reason lies in the architecture of alignment itself.
Most modern AI systems are aligned through reinforcement learning from human feedback (RLHF), where human annotators rank multiple responses and reward models learn to approximate these preferences. The language model is then optimised to maximise this learned reward signal. This process has made AI systems significantly more useful, safer, and more aligned with human expectations. But it has also introduced a structural limitation: it compresses all dimensions of human judgment into a single optimisation target.
The result is a statistical average of human preference.
And the average of everyone’s taste rarely produces memorable culture. It produces something closer to a hotel lobby painting: universally acceptable, aesthetically unobjectionable, and instantly forgettable.
Alignment systems today optimize for consensus, not taste.
A recent paper from researchers at Meta, LoRe: Personalizing LLMs via Low-Rank Reward Modeling, offers one of the clearest technical departures from this paradigm. Instead of assuming that all humans share a single reward function, LoRe proposes that preferences can be represented within a low-dimensional latent space. Rather than training one monolithic reward model for everyone, the system learns a shared basis of reward functions and represents each user as a weighted combination of these latent dimensions.
Taste becomes compositional.
The implication is profound. Human preference may feel infinitely nuanced psychologically while remaining surprisingly low-rank mathematically — much like recommendation systems that learn latent factors underlying music or film preferences. More importantly, LoRe enables few-shot personalisation: a small number of preference comparisons may be sufficient to steer generation toward an individual’s preferred style, structure, or tone.
This marks an important transition in AI alignment: from universal optimisation toward personalised alignment.
Yet even this framework leaves a deeper question unresolved. LoRe models what people prefer. It does not model why they prefer it.
Human taste is shaped not only by formal preference, but by culture, memory, identity, and historical context. Two individuals may listen to the same folk musician for entirely different reasons. One may associate the music with childhood familiarity and regional memory; another may value it as a signal of authenticity in reaction to synthetic AI-generated culture. The observable preference is identical. The underlying structure of taste is not.
A more complete theory of computational taste may therefore require at least four interacting dimensions.
The first dimension consists of formal features: the intrinsic properties of the content itself. In writing, this includes syntax, pacing, semantic density, and rhythm. In music, it includes harmony, tempo, and timbre. Contemporary language models already represent these features effectively through embeddings and large-scale pretraining.
The second dimension involves exposure pathways: how cultural artifacts reach individuals. A song discovered through a close friend carries a different psychological weight than one surfaced algorithmically. Childhood exposure, regional familiarity, and social trust all shape aesthetic attachment. Existing recommendation systems partially capture this layer, but typically flatten these pathways into static behavioral signals.
The more neglected dimensions are social positioning and temporal context. Sociologist Pierre Bourdieu argued that taste functions not merely as preference, but as cultural signaling. Preferences communicate identity, status, belonging, and distinction. The appeal of certain books, films, or musical genres often lies as much in what they signify socially as in their intrinsic content. Current reward models possess no explicit representation for this symbolic layer of taste.
Temporal context matters equally. Cultural meaning changes across time. Minimalist design may signify technological futurism in one decade and anti-technological authenticity in another. Musical genres re-emerge not because their acoustic structure objectively improves, but because they resonate with specific historical moods and collective anxieties. Preference is therefore temporally situated, not static.
A future computational model of taste would need to integrate all four layers simultaneously: formal structure, exposure pathways, social positioning, and temporal context. In such a system, taste would no longer be treated as a property of either the
user or the artifact alone. Instead, it would emerge from patterns of relationships across individuals, cultures, histories, and symbolic networks.
The implications extend far beyond recommendation systems. Educational systems, for instance, may eventually adapt explanations not only to knowledge level, but to cognitive aesthetics: some learners prefer abstraction and compression, while others rely on narrative and analogy. A computational model of taste could personalize reasoning itself.
This also introduces serious risks. Systems capable of modeling not only what individuals prefer, but why they prefer it, could shape identity and influence with unprecedented precision. The distinction between personalization and persuasion — between a system that serves human taste and one that manufactures it — may become increasingly difficult to define.
The next frontier in AI may therefore not be artificial general intelligence alone, but artificial cultural understanding. The systems that define the coming decade may not be those that simply generate the most accurate responses, but those that most effectively model the cultural structure of human taste itself.
