When an AI personalizes for someone who doesn't fit its categories, does tailoring per user, per persona, or for everyone hold up best?

This explores what happens when a personalization system meets a user or request that doesn't fit any concept map it already has, and how user-level, persona-level and global personalization each cope. No note in the collection tests this head-on. Read together, though, the notes suggest the three levels differ mainly in how much they need the concept map to be known in advance.

The three levels trade precision against scale, and each depends on the map differently (How do personalization granularity levels trade precision against scalability?). Persona-level personalization scales better but "requires domain knowledge". Someone has to have decided in advance what the persona categories are, so it is the level most exposed to an unknown ontology. User-level needs no predefined categories, but it runs into data sparsity. Global preference needs no map either, because it averages everyone together, and that erases individual differences. So the unknown gets handled by breaking (persona), by paying in data (user), or by ignoring it (global). The failure is real: personalized models lean on shallow semantic correlations and stumble when a user's profile and their query live in different concept spaces. Bridging that gap takes explicit concept-aware reasoning, not better retrieval (Why do personalized language models fail when profiles and preferences diverge?).

One response is to build the missing map instead of assuming it. An LLM can propose and refine a label taxonomy from raw text, label data with it, and distill the result into cheap classifiers (Can LLMs efficiently generate taxonomies and label training data?). That is one way to get persona-style categories without hand-written domain knowledge. A related result outside personalization shows the same move: auto-generated domain taxonomies let a model reach about half of full-corpus performance with 0.3% of the training data, because it learns where knowledge sits in a structure (Can organizing knowledge structures beat raw training data volume?). In both cases the ontology is generated rather than given.

The other response is to route around the ontology. Abstract preference summaries beat retrieving specific past interactions (Does abstract preference knowledge outperform specific interaction recall?), and profiles built only from what users wrote match complete profiles. That suggests personalization runs largely on style and preference, not subject matter (Do user outputs outperform inputs for LLM personalization?). If you model how someone writes, a concept mismatch matters less. Two learned-from-feedback approaches follow the same logic. An LLM trained with recommender feedback learns to write effective product queries without ever seeing the catalog (Can LLMs recommend products without ever seeing the catalog?). And ten adaptive questions are enough to place a user as a mix of shared base reward functions (Can user preferences be learned from just ten questions?). That is a middle path between persona and user level, where the shared categories are learned from preference data instead of named by hand. Item identifiers show a smaller version of the same hedge: numeric IDs give distinctness with no meaning attached, and titles and attributes add the meaning (Can item identifiers balance uniqueness and semantic meaning?).

The pattern is that the coarser the level, the more the ontology has to be assumed. The finer or more adaptive the level, the more it gets learned from feedback, at the cost of data or questions. There is a caution too. Personal context already pushes models toward irrelevant personal references and excessive agreement, and user profiles drove most of that shift (Does personalization make large language models worse at their jobs?). If a system guesses wrong about an unfamiliar concept space, that tendency is likely to turn the wrong guess into flattery.